mirror of
https://github.com/civitai/civitai.git
synced 2026-09-20 22:08:18 +08:00
feat(ingestion): scan images with imageScanning behind a Flipt flag
ClickUp 868m5wp5r. Image and video ingestion can submit the single
orchestrator `imageScanning` step instead of `wdTagging` + `mediaRating`,
gated by the Flipt flag `image-ingestion-image-scanning` (per image id,
off by default). Flag off, the submitted workflow is unchanged.
- `/api/webhooks/image-scan-result` stays the only callback. It fetches
the workflow and routes on its step types, so both shapes can be in
flight across a flag flip.
- Stages both pipelines share move unchanged to `image-scan-pipeline.ts`.
The legacy service keeps its own parsing and flow.
- New `image-scanning-result.service.ts` reads the imageScanning output
directly (images and video frames), keeps only general tags, and
records csam without acting on it, as legacy does.
- The new pipeline logs to Axiom as `image-scanning-result` /
`image-scanning-ingestion`, submits under
`image_scan_submitted_total{lane="imageScanning"}`, and writes scanner
audit rows as version '2'.
- Remove the non-orchestrator scanner path: the webhook's legacy body
handling, the `IMAGE_SCANNER_NEW` Redis toggle, `ingestImageBulk`,
`image.ingestArticleImages`, `/api/webhooks/reingest-images`,
`/api/internal/add-missing-phash`, `/api/mod/scan-images`, and the
`IMAGE_SCANNING_ENDPOINT` / `IMAGE_SCANNING_MODEL` env vars.
- Bump `@civitai/orchestration-client` to 0.2.0-beta.106 for the
imageScanning types.
Keep the flag off until a deploy has fully rolled out: pods on the
previous build cannot read imageScanning callbacks.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
briant
@@ -207,8 +207,6 @@ SHOPIFY_ADMIN_TOKEN=
|
||||
FLIPT_URL=""
|
||||
FLIPT_FETCHER_SECRET=placeholder
|
||||
|
||||
IMAGE_SCANNER_NEW=false
|
||||
|
||||
# App Blocks — per-app generation spend/velocity ABSOLUTE CEILINGS (incident knobs).
|
||||
# 🔴 These are UPPER BOUNDS, not the limit an app receives. Each app's actual
|
||||
# ceilings come from its server-owned `spendTier` (+ any moderator per-app
|
||||
|
||||
@@ -438,22 +438,12 @@ Both use `GREATEST` semantics (never lower, only raise) and lock `userNsfwLevel`
|
||||
|
||||
#### 4. Webhook Integration
|
||||
|
||||
**File**: `src/pages/api/webhooks/image-scan-result.ts` (image scanning)
|
||||
**File**: `src/pages/api/webhooks/image-scan-result.ts` reads the flag; the scan processors (`image-scan-result.service.ts`, `image-scanning-result.service.ts`) do the fan-out
|
||||
|
||||
```typescript
|
||||
// Feature flag gated
|
||||
const featureFlags = getFeatureFlagsLazy({ req });
|
||||
if (featureFlags.articleImageScanning) {
|
||||
// Find articles using this image
|
||||
const articleConnections = await dbWrite.imageConnection.findMany({
|
||||
where: { imageId: image.id, entityType: 'Article' },
|
||||
});
|
||||
|
||||
// Debounced updates
|
||||
for (const { entityId } of articleConnections) {
|
||||
await debounceArticleUpdate(entityId);
|
||||
}
|
||||
}
|
||||
articleImageScanning: getFeatureFlagsLazy({ req }).articleImageScanning, // webhook → processor input
|
||||
// in the processor, once the verdict is written:
|
||||
if (articleImageScanning) await fanOutArticleImageUpdates(imageId); // content + cover articles → debounceArticleUpdate
|
||||
```
|
||||
|
||||
**File**: `src/pages/api/webhooks/text-moderation-result.ts` (text moderation)
|
||||
@@ -993,7 +983,8 @@ console.log('Article Image Scanning:', flags.articleImageScanning);
|
||||
- `src/server/services/article.service.ts` - Main article operations, image linking, scan status
|
||||
- `src/server/services/article-content-cleanup.service.ts` - Image/media extraction from content (server-side)
|
||||
- `src/server/services/nsfwLevels.service.ts` - NSFW level calculation (cover + content images)
|
||||
- `src/pages/api/webhooks/image-scan-result.ts` - Image scan webhook handler
|
||||
- `src/pages/api/webhooks/image-scan-result.ts` - Image scan webhook handler (reads the flag, routes by workflow step)
|
||||
- `src/server/services/image-scan-result.service.ts`, `image-scanning-result.service.ts`, `image-scan-pipeline.ts` - scan processing + article fan-out
|
||||
- `src/server/utils/webhook-debounce.ts` - Redis-based debouncing logic
|
||||
- `src/utils/article-helpers.ts` - Image extraction (client-side), shared helpers
|
||||
|
||||
|
||||
@@ -391,7 +391,7 @@ articleImageScanning: [], // was ['public']
|
||||
```
|
||||
|
||||
**Effect**:
|
||||
- New articles still save, still submit to text moderation, but `debounceArticleUpdate` in `image-scan-result.ts` short-circuits
|
||||
- New articles still save, still submit to text moderation, but the scan processors skip `fanOutArticleImageUpdates`, so no article update is queued
|
||||
- Articles already in `Processing` will **not** auto-transition to `Published` — they need manual intervention or the flag re-enabled
|
||||
- Existing NSFW levels stay as-is
|
||||
|
||||
|
||||
@@ -187,11 +187,11 @@ one) and **Net-new** (moderation the moderator app doesn't cover — new domains
|
||||
|
||||
### 13. knights "down-leveled" review (`services/image.service.ts`)
|
||||
|
||||
- [ ] `getDownleveledImages` (ClickHouse `knights_new_order_downleveled` + PG join) + `addToNewOrderQueue`
|
||||
- [ ] `getDownleveledImages` (ClickHouse `knights_new_order_downleveled` + PG join) (`image.service.ts`) + `addToNewOrderQueue` (`image-scan-pipeline.ts`)
|
||||
|
||||
### 14. scan-result ingestion pipeline internals (`services/image-scan-result.service.ts`)
|
||||
### 14. scan-result ingestion pipeline internals (`services/image-scan-pipeline.ts`, `services/image-scan-result.service.ts`)
|
||||
|
||||
- [ ] `resolveScanOutcome`, `auditScanResults`, `markImageScanError`, `blockImageFromRating`, `getAssociatedEntities`, `evaluateImageModRules`, `isExemptFromAiVerification`, `processTags`
|
||||
- [ ] `resolveScanOutcome`, `auditScanResults`, `markImageScanError`, `getAssociatedEntities`, `evaluateImageModRules`, `processTags` (pipeline); `blockImageFromRating`, `isExemptFromAiVerification` (image-scan-result.service)
|
||||
|
||||
---
|
||||
|
||||
|
||||
@@ -43,8 +43,8 @@ ORDER BY (scanner, label, contentHash, version);
|
||||
|
||||
Column semantics:
|
||||
|
||||
- **`version`** (per-label) — policyHash from `result.policyHash` for XGuard scans; hardcoded `'1'` for image scans until the orchestrator team surfaces per-result version info on the `mediaRating` step.
|
||||
- **`modelVersion`** (workflow-level) — scanner/model version stamp, sourced from `workflow.metadata.version`. Hardcoded `'1'` everywhere today.
|
||||
- **`version`** (per-label) — policyHash from `result.policyHash` for XGuard scans; for image scans a fixed `'1'` (`mediaRating` lane) or `'2'` (`imageScanning` lane) until the orchestrator surfaces per-result version info.
|
||||
- **`modelVersion`** (workflow-level) — scanner/model version stamp, sourced from `workflow.metadata.version` (default `'1'`); image scans write `'1'` or `'2'` by lane, as above.
|
||||
|
||||
Both kept as separate columns so when the orchestrator starts returning per-label version info on `mediaRating`, we can populate `version` independently of `modelVersion`.
|
||||
|
||||
|
||||
@@ -25,7 +25,7 @@ What gets authored in phase 1:
|
||||
|
||||
The orchestrator stamps each evaluation with a `policyHash` derived from the exact policy text. That hash becomes the `version` column on `scanner_label_results`, which is what makes A/B comparison in phase 3 work without civitai-side bookkeeping.
|
||||
|
||||
For image scanners, "defining a label" means adding a new `include*` flag to the `mediaRating` step input and a corresponding output field (see [image-scan-result.service.ts](src/server/services/image-scan-result.service.ts) and the `TagSource` enum). Versioning is currently a `'1'` placeholder until the orchestrator surfaces per-result version info.
|
||||
For image scanners, "defining a label" means adding a new `include*` flag to the `mediaRating` step input and a corresponding output field (see [image-scan-result.service.ts](src/server/services/image-scan-result.service.ts) and the `TagSource` enum). Versioning is a fixed stamp until the orchestrator surfaces per-result version info: `'1'` for the `mediaRating` lane, `'2'` for the `imageScanning` lane (different models, so the two must not aggregate).
|
||||
|
||||
### Phase 2 — Provide a dataset
|
||||
|
||||
@@ -217,12 +217,12 @@ The `topK` distributions, bounding boxes, face geometry, and confidences for non
|
||||
|
||||
`wdTagging` and `mediaHash` are unchanged.
|
||||
|
||||
**Code changes required in [image-scan-result.service.ts](src/server/services/image-scan-result.service.ts):**
|
||||
**Code changes required in [image-scan-result.service.ts](src/server/services/image-scan-result.service.ts) (step types) and [image-scan-pipeline.ts](src/server/services/image-scan-pipeline.ts) (tags + audit):**
|
||||
|
||||
1. Extend the `MediaRatingStep` TypeScript type (line 74) — add the four new optional output fields.
|
||||
2. Extend `tagsWithSource` (line 251) to emit `minor`/`ai`/`anime` tags from the new fields when present.
|
||||
1. Extend the `MediaRatingStep` TypeScript type in `image-scan-result.service.ts` — add the four new optional output fields.
|
||||
2. Extend `tagsWithSource` (`buildAndInsertScanTags` in `image-scan-pipeline.ts`) to emit `minor`/`ai`/`anime` tags from the new fields when present.
|
||||
3. Add `AiRecognition` + `AnimeRecognition` to the `TagSource` Prisma enum (or accept lossy provenance and reuse `Computed`).
|
||||
4. Decide whether `auditScanResults` branches on the new `MinorDetection`-sourced `minor` tag (stronger signal than the hand-curated `tagsNeedingReview` list) or just appends `minor` to that list.
|
||||
4. Decide whether `auditScanResults` (`image-scan-pipeline.ts`) branches on the new `MinorDetection`-sourced `minor` tag (stronger signal than the hand-curated `tagsNeedingReview` list) or just appends `minor` to that list.
|
||||
5. Forward the full untruncated `mediaRating.output` to the audit log so the per-detection / topK / non-triggered confidences are preserved for tuning.
|
||||
|
||||
---
|
||||
@@ -254,13 +254,13 @@ The `topK` distributions, bounding boxes, face geometry, and confidences for non
|
||||
|
||||
### Image ingestion
|
||||
|
||||
- **Webhook**: [src/pages/api/webhooks/image-scan-result.ts](src/pages/api/webhooks/image-scan-result.ts) — handles legacy POST + new orchestrator workflow format
|
||||
- **Webhook**: [src/pages/api/webhooks/image-scan-result.ts](src/pages/api/webhooks/image-scan-result.ts) — orchestrator workflow/job events only; routes `imageScanning` workflows to [image-scanning-result.service.ts](src/server/services/image-scanning-result.service.ts) and `wdTagging` + `mediaRating` workflows to `processImageScanWorkflow`. Which step is submitted is chosen in `createImageIngestionRequest` by the Flipt flag `image-ingestion-image-scanning`.
|
||||
- **Processor**: [image-scan-result.service.ts](src/server/services/image-scan-result.service.ts) → `processImageScanWorkflow` — parses `wdTagging`/`mediaRating`/`mediaHash` steps. The `mediaRating` type definition is what needs to be extended for the new fields.
|
||||
- **Storage**:
|
||||
- `Image` table: `nsfwLevel: Int`, `minor: Boolean`, `poi: Boolean`, `needsReview: String?`, `blockedFor: String?`, `ingestion: ImageIngestionStatus`, `scanJobs: Json?`
|
||||
- `TagsOnImageDetails`: `automated`, `disabled`, `needsReview`, `confidence`, `source: TagSource`
|
||||
- `ImageTagForReview`: review queue (per-image, per-tag)
|
||||
- **Review trigger conditions** (current logic in [image-scan-result.ts](src/pages/api/webhooks/image-scan-result.ts:697-733)): `child-10/13/15` + realistic, POI word-list match, `nsfwLevel === Blocked`, moderator-specific tags.
|
||||
- **Review trigger conditions** (current logic in `auditScanResults`, [image-scan-pipeline.ts](src/server/services/image-scan-pipeline.ts)): `child-10/13/15` + realistic, POI word-list match, `nsfwLevel === Blocked`, moderator-specific tags.
|
||||
- **Tag rules**: [src/server/utils/tag-rules.ts](src/server/utils/tag-rules.ts) — replacements, appends, computed combos.
|
||||
|
||||
### Cross-cutting
|
||||
|
||||
@@ -4,6 +4,8 @@ Investigation of errors logged by the `/api/webhooks/image-scan-result` handler
|
||||
`processImageScanResult` (`src/server/services/image-scan-result.service.ts`), which became the
|
||||
default scan-result path on 2026-07-22.
|
||||
|
||||
> **2026-09 note:** `processImageScanResult` is gone. The webhook now dispatches to `processImageScanWorkflow` (wdTagging + mediaRating) or `processImageScanningWorkflow` (imageScanning), and the shared stages moved to `image-scan-pipeline.ts`. Errors log as `name: 'image-scan-result'` or `'image-scanning-result'` by lane; submit failures log as `'image-ingestion'` or `'image-scanning-ingestion'`.
|
||||
|
||||
## Where failures are logged
|
||||
|
||||
| Signal | Dataset | Shape |
|
||||
@@ -181,9 +183,9 @@ await signalClient
|
||||
This is already the house pattern — `src/server/services/referral.service.ts` and
|
||||
`src/server/auth/session-invalidation.ts` both catch-and-log around `signalClient.send`.
|
||||
|
||||
Apply the same change to the legacy path at `src/pages/api/webhooks/image-scan-result.ts` (the
|
||||
unguarded `await signalClient.send` in the `data.ingestion !== 'Blocked'` branch) — that call is
|
||||
why timeout/signal-500 errors also appear on days before the 07-22 cutover.
|
||||
The legacy (non-orchestrator) path in `src/pages/api/webhooks/image-scan-result.ts` that also sent
|
||||
this signal unguarded has since been removed; it is why timeout/signal-500 errors also appear on days
|
||||
before the 07-22 cutover.
|
||||
|
||||
### Tradeoff
|
||||
|
||||
@@ -326,5 +328,5 @@ Images rejected as "too large" end up terminally `Error` and unscanned. That is
|
||||
downscaling before submit, or raising the orchestrator's limit. Worth a follow-up decision.
|
||||
|
||||
`markImageScanSubmitFailure` duplicates the jsonb shape written by `markImageScanError` in
|
||||
`image-scan-result.service.ts`. They differ (submit failures have no `workflowId`), but the two
|
||||
`image-scan-pipeline.ts`. They differ (submit failures have no `workflowId`), but the two
|
||||
should probably share a helper before a third caller appears.
|
||||
|
||||
@@ -166,8 +166,7 @@ replica-lag miss, so a plain `dbRead` is safe.
|
||||
branches). When there's no post it does a `dbRead` lookup by `thumbnailImageId`, then calls
|
||||
`updateModel3DNsfwLevels` directly (a **synchronous** recompute — not a JobQueue enqueue).
|
||||
- Call sites (each just passes `{ imageId, postId }`, replacing the removed `queueModel3D…`):
|
||||
- `src/pages/api/webhooks/image-scan-result.ts` — Scanned + Blocked
|
||||
- `src/server/services/image-scan-result.service.ts` — Scanned + Blocked branches
|
||||
- `src/server/services/image-scan-pipeline.ts` `applyIngestionSideEffects` — Scanned + Blocked branches (shared by both scan pipelines; the legacy webhook body is gone)
|
||||
- `src/server/services/image.service.ts` `updateImageNsfwLevel` — moderator branch (also
|
||||
added `postId` to the `findUnique` select)
|
||||
|
||||
@@ -195,7 +194,7 @@ Model3D trigger) and had the image trigger read it (zero lookup). A review found
|
||||
soundness gap: `Image.metadata` is rewritten **wholesale** in several places, so the flag
|
||||
could be silently and permanently dropped:
|
||||
|
||||
- `src/server/services/image-scan-result.service.ts` (~line 735): `"metadata" =
|
||||
- `src/server/services/image-scan-pipeline.ts` (`resolveScanOutcome`): `"metadata" =
|
||||
COALESCE(<modRule.metadata>::jsonb, "metadata")` **replaces** the column when a mod rule
|
||||
carries a metadata payload — on the scan path, same UPDATE as the nsfwLevel change.
|
||||
- `src/server/services/image.service.ts` `updateImageNsfwLevel` (~line 6747) and the
|
||||
|
||||
@@ -111,7 +111,7 @@ Re-read off the live ticket 2026-08-21; the section above had captured the middl
|
||||
every gallery image via `UPDATE "Image" SET minor = …, poi = …`, queueing those ids into the image
|
||||
index. `poi` is on `modelUpsertSchema` beside `minor` and `upsertModel` calls that fan-out, so
|
||||
flipping poi today already does all of it. Image-level `poi` also excludes from the image search
|
||||
index and is re-derived from the parent model at scan time (`image-scan-result.service.ts:1089`).
|
||||
index and is re-derived from the parent model at scan time (`getAssociatedEntities` in `image-scan-pipeline.ts`).
|
||||
|
||||
What `poi` genuinely lacks against `minor` is narrower, and is a build list rather than a question:
|
||||
**no dedicated setter / `moderatorProcedure`**, **no snapshot** (so no rollback), **absent from
|
||||
|
||||
@@ -170,7 +170,8 @@ and builds; that is all that is established. Before any of it is trusted in prod
|
||||
## 4. Known-open, decided or deferred
|
||||
|
||||
- **`aiNsfwLevel` / `aiModel` exist in production but not in `schema.full.prisma`.**
|
||||
`src/pages/api/webhooks/image-scan-result.ts` writes both, and Front Page Audit reads `aiNsfwLevel`
|
||||
nothing in this repo writes them any more (the only writer was the legacy scanner branch of
|
||||
`image-scan-result.ts`, removed), and Front Page Audit reads `aiNsfwLevel`
|
||||
through raw `sql` because it cannot be selected as a typed column. It is the scanner's own rating, and
|
||||
disagreement with `nsfwLevel` is the strongest signal that a row needs a human — so it is worth
|
||||
having. Either add both to the schema and regenerate, or accept the raw read as permanent.
|
||||
|
||||
@@ -118,7 +118,6 @@ const IO_CALL_NAMES = new Set([
|
||||
'fetch',
|
||||
// image ingestion / scanner
|
||||
'ingestImage',
|
||||
'ingestImageBulk',
|
||||
'createImageIngestionRequest',
|
||||
// orchestrator
|
||||
'submitWorkflow',
|
||||
|
||||
+1
-1
@@ -140,7 +140,7 @@
|
||||
"@civitai/generation-metadata": "^0.3.0",
|
||||
"@civitai/moderation": "workspace:*",
|
||||
"@civitai/next-axiom": "^0.17.0",
|
||||
"@civitai/orchestration-client": "0.2.0-beta.104",
|
||||
"@civitai/orchestration-client": "0.2.0-beta.106",
|
||||
"@civitai/shared": "workspace:*",
|
||||
"@clavata/sdk": "^0.2.3",
|
||||
"@clickhouse/client": "^0.2.2",
|
||||
|
||||
@@ -2064,15 +2064,6 @@ export const REDIS_SYS_KEYS = {
|
||||
CREATION_BLOCKED_TAGS: 'system:creation-blocked-tags',
|
||||
LIVE_FEATURE_FLAGS: 'system:live-feature-flags',
|
||||
SUSPICIOUS_AUDIT_MATCHES: 'system:suspicious-audit-matches',
|
||||
/*
|
||||
Runtime toggle for the new image ingestion path (createImageIngestionRequest
|
||||
with the expanded mediaRating step). Read by image.service.ts before
|
||||
routing to the new vs legacy scanner. Accepts '1'/'true' to enable,
|
||||
'0'/'false' to disable. If the key is missing, the first call seeds it to
|
||||
'false' so the toggle is discoverable in Redis. Lets ops flip without a
|
||||
deploy.
|
||||
*/
|
||||
IMAGE_SCANNER_NEW: 'system:image-scanner-new',
|
||||
/*
|
||||
Per-run image cap for the remove-deleted-user-images job. Set to '0' to
|
||||
pause the drain without a deploy. Missing key means the job's compiled
|
||||
|
||||
@@ -586,7 +586,7 @@ export const imageScanWebhookCounter = registerCounterWithLabels({
|
||||
|
||||
export const imageScanSubmittedCounter = registerCounterWithLabels({
|
||||
name: 'image_scan_submitted_total',
|
||||
help: 'ingestImage() scan submissions by lane (new|legacy) and result (success|failed)',
|
||||
help: 'ingestImage() scan submissions by lane (new = wdTagging+mediaRating | imageScanning) and result (success|failed)',
|
||||
labelNames: ['lane', 'result'] as const,
|
||||
});
|
||||
|
||||
|
||||
Generated
+5
-5
@@ -71,8 +71,8 @@ importers:
|
||||
specifier: ^0.17.0
|
||||
version: 0.17.0(next@16.3.1(@babel/core@7.29.0)(@opentelemetry/api@1.9.0)(@playwright/test@1.57.0)(@types/node@24.13.3)(react-dom@18.3.1(react@18.3.1))(react@18.3.1)(sass@1.90.0))
|
||||
'@civitai/orchestration-client':
|
||||
specifier: 0.2.0-beta.104
|
||||
version: 0.2.0-beta.104
|
||||
specifier: 0.2.0-beta.106
|
||||
version: 0.2.0-beta.106
|
||||
'@civitai/shared':
|
||||
specifier: workspace:*
|
||||
version: link:packages/civitai-shared
|
||||
@@ -2225,8 +2225,8 @@ packages:
|
||||
peerDependencies:
|
||||
next: ^12.1.4 || ^13 || ^14
|
||||
|
||||
'@civitai/orchestration-client@0.2.0-beta.104':
|
||||
resolution: {integrity: sha512-V7EKGBMi3YjtOfkxXOmogJhSbM83t9OCJOtxH6QSpwzsE7JZH85gtoMW6sSvF7pDLj9Ywim68Pmlq2M0Qd47sg==}
|
||||
'@civitai/orchestration-client@0.2.0-beta.106':
|
||||
resolution: {integrity: sha512-PfDmcOXhwgknRitkyeE21/oJMnu6+eDSYyLWL2UgdZScoi5adRnEhJNfV+nWIBHzKz62whwEW+BhG8uxTTzsZw==}
|
||||
engines: {git: '>=2.11.0', node: '>=18.0.0', npm: '>=7.19.0', yarn: '>=1.7.0'}
|
||||
|
||||
'@clavata/sdk@0.2.3':
|
||||
@@ -13450,7 +13450,7 @@ snapshots:
|
||||
next: 16.3.1(@babel/core@7.29.0)(@opentelemetry/api@1.9.0)(@playwright/test@1.57.0)(@types/node@24.13.3)(react-dom@18.3.1(react@18.3.1))(react@18.3.1)(sass@1.90.0)
|
||||
whatwg-fetch: 3.6.20
|
||||
|
||||
'@civitai/orchestration-client@0.2.0-beta.104':
|
||||
'@civitai/orchestration-client@0.2.0-beta.106':
|
||||
dependencies:
|
||||
'@hey-api/client-fetch': 0.1.14
|
||||
rfc6902: 5.2.0
|
||||
|
||||
@@ -0,0 +1,142 @@
|
||||
import type { NextApiRequest, NextApiResponse } from 'next';
|
||||
import { beforeEach, describe, expect, it, vi } from 'vitest';
|
||||
import type * as ImageScanningResultService from '~/server/services/image-scanning-result.service';
|
||||
|
||||
const mocks = vi.hoisted(() => ({
|
||||
getWorkflow: vi.fn(),
|
||||
captureJobFailureReason: vi.fn(),
|
||||
processImageScanWorkflow: vi.fn(),
|
||||
processImageScanningWorkflow: vi.fn(),
|
||||
}));
|
||||
|
||||
vi.mock('~/server/utils/endpoint-helpers', () => ({
|
||||
WebhookEndpoint: (handler: unknown) => handler,
|
||||
}));
|
||||
vi.mock('@civitai/client', () => ({ getWorkflow: mocks.getWorkflow }));
|
||||
vi.mock('~/server/services/orchestrator/client', () => ({ internalOrchestratorClient: {} }));
|
||||
vi.mock('~/server/services/feature-flags.service', () => ({
|
||||
getFeatureFlagsLazy: () => ({ articleImageScanning: true }),
|
||||
}));
|
||||
vi.mock('~/server/prom/client', () => ({ imageScanWebhookCounter: { inc: vi.fn() } }));
|
||||
// Both pipelines import the whole ingestion stack; this suite pins only the routing between them.
|
||||
vi.mock('~/server/services/image-scan-pipeline', () => ({
|
||||
captureJobFailureReason: mocks.captureJobFailureReason,
|
||||
LEGACY_SCAN_LOG: { name: 'image-scan-result', source: 'image-scan-result.service' },
|
||||
}));
|
||||
vi.mock('~/server/services/image-scan-result.service', () => ({
|
||||
processImageScanWorkflow: mocks.processImageScanWorkflow,
|
||||
}));
|
||||
// The real `isImageScanningWorkflow` makes the routing decision under test.
|
||||
vi.mock('~/server/services/image-scanning-result.service', async (importOriginal) => ({
|
||||
...(await importOriginal<typeof ImageScanningResultService>()),
|
||||
processImageScanningWorkflow: mocks.processImageScanningWorkflow,
|
||||
}));
|
||||
vi.mock('~/server/services/scanner-audit.service', () => ({ recordImageScanningResult: vi.fn() }));
|
||||
vi.mock('~/server/services/job-queue.service', () => ({ removeImageScanJobQueue: vi.fn() }));
|
||||
vi.mock('~/server/services/orchestrator/orchestrator.service', () => ({
|
||||
computePerceptualHash: vi.fn(),
|
||||
}));
|
||||
vi.mock('~/server/utils/webhook-debounce', () => ({ fanOutArticleImageUpdates: vi.fn() }));
|
||||
|
||||
import handler from '~/pages/api/webhooks/image-scan-result';
|
||||
import { loggingMock } from '~/__tests__/mocks/logging.mock';
|
||||
|
||||
const call = async (body: unknown) => {
|
||||
const res = {
|
||||
status: vi.fn().mockReturnThis(),
|
||||
json: vi.fn().mockReturnThis(),
|
||||
send: vi.fn().mockReturnThis(),
|
||||
};
|
||||
await (handler as unknown as (req: NextApiRequest, res: NextApiResponse) => Promise<unknown>)(
|
||||
{ method: 'POST', body } as NextApiRequest,
|
||||
res as unknown as NextApiResponse
|
||||
);
|
||||
return res;
|
||||
};
|
||||
const deliver = (status: string, steps: unknown[]) => {
|
||||
mocks.getWorkflow.mockResolvedValue({
|
||||
data: { metadata: { imageId: 7 }, startedAt: 'start', completedAt: 'end', steps },
|
||||
});
|
||||
return call({ workflowId: 'wf', status });
|
||||
};
|
||||
const pipelineInput = (status: string, steps: unknown[]) => ({
|
||||
workflowId: 'wf',
|
||||
status,
|
||||
steps,
|
||||
imageId: 7,
|
||||
articleImageScanning: true,
|
||||
startedAt: 'start',
|
||||
completedAt: 'end',
|
||||
});
|
||||
|
||||
describe('image-scan-result webhook routing', () => {
|
||||
beforeEach(() => vi.clearAllMocks());
|
||||
|
||||
it('stashes a job failure reason without fetching the workflow', async () => {
|
||||
const event = { workflowId: 'wf', jobId: 'job', reason: 'timed out' };
|
||||
const res = await call(event);
|
||||
|
||||
expect(mocks.captureJobFailureReason).toHaveBeenCalledWith(event);
|
||||
expect(mocks.getWorkflow).not.toHaveBeenCalled();
|
||||
expect(res.status).toHaveBeenCalledWith(200);
|
||||
});
|
||||
|
||||
it('sends an imageScanning workflow to the imageScanning pipeline, failed or not', async () => {
|
||||
const steps = [{ $type: 'imageScanning', status: 'failed' }];
|
||||
await deliver('failed', steps);
|
||||
|
||||
expect(mocks.processImageScanningWorkflow).toHaveBeenCalledTimes(1);
|
||||
expect(mocks.processImageScanningWorkflow).toHaveBeenCalledWith(pipelineInput('failed', steps));
|
||||
expect(mocks.processImageScanWorkflow).not.toHaveBeenCalled();
|
||||
});
|
||||
|
||||
it('sends a wdTagging + mediaRating workflow to the legacy pipeline', async () => {
|
||||
const steps = [
|
||||
{ $type: 'wdTagging', output: { tags: {} } },
|
||||
{ $type: 'mediaRating', output: { nsfwLevel: 'pg' } },
|
||||
];
|
||||
await deliver('succeeded', steps);
|
||||
|
||||
expect(mocks.processImageScanWorkflow).toHaveBeenCalledTimes(1);
|
||||
expect(mocks.processImageScanWorkflow).toHaveBeenCalledWith(pipelineInput('succeeded', steps));
|
||||
expect(mocks.processImageScanningWorkflow).not.toHaveBeenCalled();
|
||||
});
|
||||
|
||||
it('rejects a workflow it cannot load or that carries no imageId', async () => {
|
||||
mocks.getWorkflow.mockResolvedValueOnce({ data: undefined });
|
||||
const missing = await call({ workflowId: 'wf', status: 'succeeded' });
|
||||
expect(missing.status).toHaveBeenCalledWith(400);
|
||||
expect(missing.send).toHaveBeenCalledWith({ error: 'could not find workflow: wf' });
|
||||
|
||||
mocks.getWorkflow.mockResolvedValueOnce({ data: { metadata: {}, steps: [] } });
|
||||
const noImage = await call({ workflowId: 'wf', status: 'succeeded' });
|
||||
expect(noImage.send).toHaveBeenCalledWith({
|
||||
error: 'missing workflow metadata.imageId - wf',
|
||||
});
|
||||
});
|
||||
|
||||
it.each([
|
||||
['imageScanning', [{ $type: 'imageScanning' }], 'image-scanning-result'],
|
||||
['legacy', [{ $type: 'wdTagging' }, { $type: 'mediaRating' }], 'image-scan-result'],
|
||||
])('logs a %s pipeline failure under its own name', async (_, steps, name) => {
|
||||
const pipeline =
|
||||
name === 'image-scanning-result'
|
||||
? mocks.processImageScanningWorkflow
|
||||
: mocks.processImageScanWorkflow;
|
||||
pipeline.mockRejectedValueOnce(new Error('boom'));
|
||||
const res = await deliver('succeeded', steps);
|
||||
|
||||
expect(res.status).toHaveBeenCalledWith(400);
|
||||
expect(loggingMock.logToAxiom).toHaveBeenCalledWith(
|
||||
expect.objectContaining({ name, type: 'error', message: 'boom' })
|
||||
);
|
||||
});
|
||||
|
||||
it('acknowledges a result for an image deleted since the scan was submitted', async () => {
|
||||
mocks.processImageScanningWorkflow.mockRejectedValueOnce(new Error('image not found: 7'));
|
||||
const res = await deliver('succeeded', [{ $type: 'imageScanning', output: {} }]);
|
||||
|
||||
expect(res.status).toHaveBeenCalledWith(200);
|
||||
expect(res.json).toHaveBeenCalledWith({ ok: true, skipped: 'deleted' });
|
||||
});
|
||||
});
|
||||
@@ -1,11 +1,10 @@
|
||||
import { describe, it, expect, vi, beforeEach } from 'vitest';
|
||||
import type { NextApiRequest, NextApiResponse } from 'next';
|
||||
import handler from '~/pages/api/webhooks/image-scan-result';
|
||||
import { processImageScanWorkflow } from '~/server/services/image-scan-result.service';
|
||||
import { processImageScanningWorkflow } from '~/server/services/image-scanning-result.service';
|
||||
import { clickhouse } from '~/server/clickhouse/client';
|
||||
import { signalClient } from '~/utils/signal-client';
|
||||
import type * as ClickhouseClient from '~/server/clickhouse/client';
|
||||
import { TagSource, ImageIngestionStatus } from '~/shared/utils/prisma/enums';
|
||||
import { NsfwLevel } from '~/server/common/enums';
|
||||
import type * as BlocklistService from '~/server/services/blocklist.service';
|
||||
|
||||
const {
|
||||
@@ -271,7 +270,6 @@ vi.mock('~/server/services/image.service', () => ({
|
||||
getImagesModRules: vi.fn().mockResolvedValue([]),
|
||||
queueImageSearchIndexUpdate: vi.fn().mockResolvedValue(undefined),
|
||||
enqueueImageIngestion: vi.fn().mockResolvedValue(undefined),
|
||||
imageScanTypes: [3, 9], // ImageScanType.WD14, ImageScanType.SpineRating
|
||||
}));
|
||||
|
||||
// Render a Prisma tagged-template call the way the driver does: nested `Prisma.sql` fragments
|
||||
@@ -403,158 +401,6 @@ describe('image-scan-result webhook - pipeline tests', () => {
|
||||
});
|
||||
});
|
||||
|
||||
const runWebhook = (body: any) => {
|
||||
const req = {
|
||||
method: 'POST',
|
||||
query: { token: 'mock-webhook-token' },
|
||||
headers: { host: 'localhost:3000' },
|
||||
body,
|
||||
} as unknown as NextApiRequest;
|
||||
|
||||
const res = {
|
||||
status: vi.fn().mockReturnThis(),
|
||||
json: vi.fn().mockImplementation((val) => res),
|
||||
send: vi.fn().mockImplementation((val) => res),
|
||||
} as unknown as NextApiResponse;
|
||||
|
||||
return { promise: handler(req, res), res };
|
||||
};
|
||||
|
||||
it('should prevent tag-bleed between Clavata (ignored) and WD14 (not ignored) concurrently', async () => {
|
||||
const reqA = runWebhook({
|
||||
id: 1,
|
||||
status: 0,
|
||||
source: TagSource.Clavata,
|
||||
tags: [{ tag: 'hate symbols', confidence: 95 }],
|
||||
});
|
||||
|
||||
const reqB = runWebhook({
|
||||
id: 2,
|
||||
status: 0,
|
||||
source: TagSource.WD14,
|
||||
tags: [{ tag: 'hate symbols', confidence: 95 }],
|
||||
});
|
||||
|
||||
await Promise.all([reqA.promise, reqB.promise]);
|
||||
|
||||
expect(reqA.res.status).toHaveBeenCalledWith(200);
|
||||
expect(reqB.res.status).toHaveBeenCalledWith(200);
|
||||
|
||||
const dbUpdates = mockDbWrite.image.update.mock.calls;
|
||||
const updateForImage2 = dbUpdates.find((call: any) => call[0].where.id === 2);
|
||||
|
||||
expect(updateForImage2).toBeDefined();
|
||||
expect(updateForImage2[0].data.nsfwLevel).toBe(32); // Blocked
|
||||
expect(updateForImage2[0].data.ingestion).toBe(ImageIngestionStatus.Scanned);
|
||||
|
||||
const insertedTags = [
|
||||
...mockInsertTagsOnImageNew.mock.calls.flatMap((call) => call[0]),
|
||||
...mockUpsertTagsOnImageNew.mock.calls.flatMap((call) => call[0]),
|
||||
];
|
||||
const tagForImage1 = insertedTags.find((t) => t.imageId === 1 && t.tagId === 100);
|
||||
const tagForImage2 = insertedTags.find((t) => t.imageId === 2 && t.tagId === 100);
|
||||
|
||||
expect(tagForImage1).toBeDefined();
|
||||
expect(tagForImage1.disabled).toBe(true); // Clavata ignores hate symbols
|
||||
|
||||
expect(tagForImage2).toBeDefined();
|
||||
expect(tagForImage2.disabled).toBe(false); // WD14 does not ignore hate symbols
|
||||
});
|
||||
|
||||
it('should set ingestion to NotFound when status is NotFound', async () => {
|
||||
const req = runWebhook({
|
||||
id: 3,
|
||||
status: 1, // NotFound
|
||||
source: TagSource.WD14,
|
||||
tags: [],
|
||||
});
|
||||
|
||||
await req.promise;
|
||||
|
||||
expect(req.res.status).toHaveBeenCalledWith(200);
|
||||
expect(mockDbWrite.image.updateMany).toHaveBeenCalledWith({
|
||||
where: { id: 3, ingestion: { in: ['Pending', 'Error'] } },
|
||||
data: { ingestion: ImageIngestionStatus.NotFound },
|
||||
});
|
||||
});
|
||||
|
||||
it('should increment retryCount when status is Unscannable', async () => {
|
||||
const req = runWebhook({
|
||||
id: 4,
|
||||
status: 2, // Unscannable
|
||||
source: TagSource.WD14,
|
||||
tags: [],
|
||||
});
|
||||
|
||||
await req.promise;
|
||||
|
||||
expect(req.res.status).toHaveBeenCalledWith(200);
|
||||
const queryCall = imageUpdates.find(
|
||||
(update) => update.text.includes('retryCount') && update.params.includes(4)
|
||||
);
|
||||
expect(queryCall).toBeDefined();
|
||||
});
|
||||
|
||||
it('should set needsReview: minor when minor tag is present and image is NSFW', async () => {
|
||||
const req = runWebhook({
|
||||
id: 5,
|
||||
status: 0,
|
||||
source: TagSource.WD14,
|
||||
tags: [
|
||||
{ tag: 'teen', confidence: 95 },
|
||||
{ tag: 'r', confidence: 95 },
|
||||
],
|
||||
});
|
||||
|
||||
await req.promise;
|
||||
|
||||
expect(req.res.status).toHaveBeenCalledWith(200);
|
||||
const dbUpdates = mockDbWrite.image.update.mock.calls;
|
||||
const updateForImage5 = dbUpdates.find((call: any) => call[0].where.id === 5);
|
||||
expect(updateForImage5).toBeDefined();
|
||||
expect(updateForImage5[0].data.needsReview).toBe('minor');
|
||||
});
|
||||
|
||||
it('should set needsReview: poi when POI tag is present', async () => {
|
||||
const req = runWebhook({
|
||||
id: 6,
|
||||
status: 0,
|
||||
source: TagSource.WD14,
|
||||
tags: [
|
||||
{ tag: 'potential celebrity', confidence: 95 },
|
||||
{ tag: 'pg', confidence: 95 },
|
||||
],
|
||||
});
|
||||
|
||||
await req.promise;
|
||||
|
||||
expect(req.res.status).toHaveBeenCalledWith(200);
|
||||
const dbUpdates = mockDbWrite.image.update.mock.calls;
|
||||
const updateForImage6 = dbUpdates.find((call: any) => call[0].where.id === 6);
|
||||
expect(updateForImage6).toBeDefined();
|
||||
expect(updateForImage6[0].data.needsReview).toBe('poi');
|
||||
});
|
||||
|
||||
it('should not update nsfwLevel if nsfwLevelLocked is true', async () => {
|
||||
const req = runWebhook({
|
||||
id: 7,
|
||||
status: 0,
|
||||
source: TagSource.WD14,
|
||||
tags: [
|
||||
{ tag: 'some-tag', confidence: 95 },
|
||||
{ tag: 'x', confidence: 95 },
|
||||
],
|
||||
});
|
||||
|
||||
await req.promise;
|
||||
|
||||
expect(req.res.status).toHaveBeenCalledWith(200);
|
||||
const dbUpdates = mockDbWrite.image.update.mock.calls;
|
||||
const updateForImage7 = dbUpdates.find((call: any) => call[0].where.id === 7);
|
||||
expect(updateForImage7).toBeDefined();
|
||||
expect(updateForImage7[0].data.nsfwLevel).toBeUndefined(); // locked, so undefined (not updated)
|
||||
});
|
||||
|
||||
it('creates a never-before-seen tag with the columns the Tag table requires', async () => {
|
||||
const passthrough = mockDbWrite.$queryRaw.getMockImplementation();
|
||||
let tagInsert: { text: string; params: any[] } | undefined;
|
||||
@@ -746,208 +592,202 @@ describe('image-scan-result webhook - pipeline tests', () => {
|
||||
expect(errorFlips).toHaveLength(0);
|
||||
});
|
||||
|
||||
describe('webhook body strings never reach SQL text', () => {
|
||||
const INJECTED = `1 OR 1=1) < 5 AND disabled = false -- `;
|
||||
|
||||
it('rejects a non-numeric hash before it can reach the ClickHouse query', async () => {
|
||||
envOverrides.BLOCKED_IMAGE_HASH_CHECK = true;
|
||||
|
||||
const req = runWebhook({
|
||||
id: 8,
|
||||
status: 0,
|
||||
source: TagSource.ImageHash,
|
||||
hash: INJECTED,
|
||||
});
|
||||
await req.promise;
|
||||
|
||||
expect(mockClickhouseQuery).not.toHaveBeenCalled();
|
||||
expect(req.res.status).toHaveBeenCalledWith(400);
|
||||
// The imageScanning pipeline through the real shared stages; only the database, ClickHouse
|
||||
// and signals are faked. Scan outputs follow the shape the orchestrator returned on 2026-09-17.
|
||||
describe('imageScanning workflows through the shared pipeline', () => {
|
||||
const scanOutput = (overrides: Record<string, unknown> = {}) => ({
|
||||
nsfwLevel: 'x',
|
||||
score: 0.91,
|
||||
topK: [{ label: 'X', score: 0.91 }],
|
||||
aiRecognition: { label: 'AI', score: 0.86, topK: [] },
|
||||
animeRecognition: { label: 'anime', score: 0.9, scores: {} },
|
||||
humanRecognition: {
|
||||
status: 'ok',
|
||||
ran: true,
|
||||
label: 'human',
|
||||
score: 0.7,
|
||||
humanScore: 0.7,
|
||||
noHumanScore: 0.3,
|
||||
scores: {},
|
||||
evidence: [],
|
||||
},
|
||||
tagging: {
|
||||
status: 'ran',
|
||||
ran: true,
|
||||
threshold: 0.55,
|
||||
tagCount: 3,
|
||||
totalAboveThreshold: 3,
|
||||
truncated: false,
|
||||
tags: [
|
||||
{ tag: 'teen', category: 'general', score: 0.9 },
|
||||
{ tag: 'hate symbols', category: 'copyright', score: 0.8 },
|
||||
{ tag: 'some-tag', category: 'meta', score: 0.8 },
|
||||
],
|
||||
},
|
||||
jointAgeClassification: {
|
||||
status: 'ran',
|
||||
ran: true,
|
||||
detections: [{ ageBand: '21-24', under18Probability: 0.06, isMinor: false }],
|
||||
minorDetected: false,
|
||||
},
|
||||
csam: false,
|
||||
...overrides,
|
||||
});
|
||||
|
||||
it('logs a match against the parsed hash, not the string the scanner sent', async () => {
|
||||
envOverrides.BLOCKED_IMAGE_HASH_CHECK = true;
|
||||
mockClickhouseQuery.mockResolvedValue([{ count: 1 }]);
|
||||
|
||||
const req = runWebhook({ id: 16, status: 0, source: TagSource.ImageHash, hash: ' 42 ' });
|
||||
await req.promise;
|
||||
|
||||
const match = mockLogToAxiom.mock.calls.find(
|
||||
(call) => call[0]?.message === 'Image pHash matched a blocked image'
|
||||
);
|
||||
expect(match).toBeDefined();
|
||||
expect(match![0].pHash).toBe('42');
|
||||
const scanStep = (output: unknown, status = 'succeeded') => ({
|
||||
$type: 'imageScanning',
|
||||
name: 'scan',
|
||||
status,
|
||||
output,
|
||||
});
|
||||
|
||||
it('passes a valid hash to ClickHouse as digits only', async () => {
|
||||
envOverrides.BLOCKED_IMAGE_HASH_CHECK = true;
|
||||
|
||||
const req = runWebhook({
|
||||
id: 9,
|
||||
status: 0,
|
||||
source: TagSource.ImageHash,
|
||||
hash: '-1234567890123456789',
|
||||
});
|
||||
await req.promise;
|
||||
|
||||
expect(mockClickhouseQuery).toHaveBeenCalled();
|
||||
const [strings, ...values] = mockClickhouseQuery.mock.calls[0];
|
||||
expect(strings.join('?')).toContain('bitXor(hash, ?)');
|
||||
expect(values).toEqual([-1234567890123456789n]);
|
||||
});
|
||||
|
||||
it('sends the scanner-supplied model id as a parameter, not as SQL text', async () => {
|
||||
const req = runWebhook({
|
||||
id: 10,
|
||||
status: 0,
|
||||
source: TagSource.WD14,
|
||||
tags: [],
|
||||
context: { movie_rating: 'PG', movie_rating_model_id: `x'; DROP TABLE "Image"; --` },
|
||||
});
|
||||
await req.promise;
|
||||
|
||||
const update = imageUpdates.find((u) => u.text.includes('"aiModel"'));
|
||||
expect(update).toBeDefined();
|
||||
expect(update!.text).not.toContain('DROP TABLE');
|
||||
expect(update!.params).toContain(`x'; DROP TABLE "Image"; --`);
|
||||
});
|
||||
});
|
||||
|
||||
describe('perceptual hash handling', () => {
|
||||
it('records the scan without a lookup when the hash is blank', async () => {
|
||||
envOverrides.BLOCKED_IMAGE_HASH_CHECK = true;
|
||||
|
||||
const req = runWebhook({ id: 11, status: 0, source: TagSource.ImageHash, hash: ' ' });
|
||||
await req.promise;
|
||||
|
||||
expect(mockClickhouseQuery).not.toHaveBeenCalled();
|
||||
expect(req.res.status).toHaveBeenCalledWith(200);
|
||||
const update = imageUpdates.find((u) => u.text.includes('jsonb_build_object'));
|
||||
expect(update).toBeDefined();
|
||||
expect(update!.text).not.toContain('"pHash"');
|
||||
expect(
|
||||
mockLogToAxiom.mock.calls.some(
|
||||
(call) => call[0]?.message === 'blank hash from ImageHash scan'
|
||||
)
|
||||
).toBe(true);
|
||||
});
|
||||
|
||||
it('stores a zero hash but does not look it up', async () => {
|
||||
envOverrides.BLOCKED_IMAGE_HASH_CHECK = true;
|
||||
|
||||
const req = runWebhook({ id: 15, status: 0, source: TagSource.ImageHash, hash: '0' });
|
||||
await req.promise;
|
||||
|
||||
const update = imageUpdates.find((u) => u.text.includes('"pHash"'));
|
||||
expect(update).toBeDefined();
|
||||
expect(update!.params).toContain(0n);
|
||||
expect(mockClickhouseQuery).not.toHaveBeenCalled();
|
||||
});
|
||||
|
||||
it('completes the scan when the blocklist lookup fails', async () => {
|
||||
envOverrides.BLOCKED_IMAGE_HASH_CHECK = true;
|
||||
mockClickhouseQuery.mockRejectedValue(new Error('socket hang up'));
|
||||
|
||||
const req = runWebhook({ id: 14, status: 0, source: TagSource.ImageHash, hash: '42' });
|
||||
await req.promise;
|
||||
|
||||
expect(mockClickhouseQuery).toHaveBeenCalled();
|
||||
expect(req.res.status).toHaveBeenCalledWith(200);
|
||||
const update = imageUpdates.find((u) => u.params.includes(42n));
|
||||
expect(update).toBeDefined();
|
||||
});
|
||||
|
||||
it('rejects a hash wider than Int64 rather than letting ClickHouse reject it', async () => {
|
||||
envOverrides.BLOCKED_IMAGE_HASH_CHECK = true;
|
||||
|
||||
const req = runWebhook({
|
||||
id: 12,
|
||||
status: 0,
|
||||
source: TagSource.ImageHash,
|
||||
hash: '99999999999999999999',
|
||||
});
|
||||
await req.promise;
|
||||
|
||||
expect(mockClickhouseQuery).not.toHaveBeenCalled();
|
||||
expect(req.res.status).toHaveBeenCalledWith(400);
|
||||
});
|
||||
|
||||
it('binds a hash as a bigint parameter, not as SQL text', async () => {
|
||||
const req = runWebhook({
|
||||
id: 13,
|
||||
status: 0,
|
||||
source: TagSource.ImageHash,
|
||||
hash: '4611686018427387904',
|
||||
});
|
||||
await req.promise;
|
||||
|
||||
const update = imageUpdates.find((u) => u.text.includes('"pHash"'));
|
||||
expect(update).toBeDefined();
|
||||
expect(update!.params).toContain(4611686018427387904n);
|
||||
expect(update!.text).not.toContain('4611686018427387904');
|
||||
});
|
||||
});
|
||||
|
||||
describe('moderator benign phrases reach the POI tagger', () => {
|
||||
// The file's own `beforeEach` is `vi.clearAllMocks()`, which clears calls but NOT
|
||||
// implementations — so the one a test below installs would leak into anything appended
|
||||
// after this describe. Harmless while this is the last block, which is exactly the kind
|
||||
// of "harmless" that stops being true without anyone noticing.
|
||||
beforeEach(() => {
|
||||
mockStripBenignPhrases.mockImplementation(async (text?: string) => text ?? '');
|
||||
});
|
||||
|
||||
const seedImageWithPrompt = (id: number, prompt: string) => {
|
||||
imageDbState.set(id, {
|
||||
id,
|
||||
createdAt: new Date(),
|
||||
scannedAt: null,
|
||||
type: 'image',
|
||||
userId: 1,
|
||||
meta: { prompt },
|
||||
metadata: {},
|
||||
postId: null,
|
||||
nsfwLevelLocked: false,
|
||||
nsfwLevel: null,
|
||||
scanJobs: { scans: {} },
|
||||
ingestion: ImageIngestionStatus.Pending,
|
||||
});
|
||||
const hashStep = {
|
||||
$type: 'mediaHash',
|
||||
name: 'hash',
|
||||
status: 'succeeded',
|
||||
output: { hashes: { perceptual: '6F51B11C49611E0E' } },
|
||||
};
|
||||
const deliver = (imageId: number, steps: unknown[], status = 'succeeded') =>
|
||||
processImageScanningWorkflow({ workflowId: `wf-${imageId}`, status, steps, imageId });
|
||||
|
||||
const requestedTagNames = () =>
|
||||
mockDbWrite.tag.findMany.mock.calls.flatMap((call: any) => call[0]?.where?.name?.in ?? []);
|
||||
const sqlOf = (call: any[]) =>
|
||||
renderSql(Array.isArray(call[0]) ? call[0] : call[0].strings, call.slice(1));
|
||||
// The verdict row written by resolveScanOutcome.
|
||||
const verdictFor = (imageId: number) => {
|
||||
const update = mockDbWrite.$executeRaw.mock.calls
|
||||
.map(sqlOf)
|
||||
.find(
|
||||
(sql: any) =>
|
||||
sql.text.includes('UPDATE "Image"') &&
|
||||
sql.text.includes('"needsReview"') &&
|
||||
sql.params.at(-1) === imageId
|
||||
);
|
||||
if (!update) return undefined;
|
||||
const param = (column: string) => {
|
||||
const match = update.text.match(new RegExp(`"${column}" = \\$(\\d+)`));
|
||||
return match ? update.params[Number(match[1]) - 1] : undefined;
|
||||
};
|
||||
return {
|
||||
ingestion: param('ingestion'),
|
||||
nsfwLevel: param('nsfwLevel'),
|
||||
needsReview: param('needsReview'),
|
||||
minor: param('minor'),
|
||||
pHash: param('pHash'),
|
||||
};
|
||||
};
|
||||
const tagsWritten = (imageId: number) =>
|
||||
mockInsertTagsOnImageNew.mock.calls
|
||||
.flatMap((call) => call[0])
|
||||
.filter((tag: any) => tag.imageId === imageId);
|
||||
const errorFlipsFor = (imageId: number) =>
|
||||
mockDbWrite.$queryRaw.mock.calls
|
||||
.map(sqlOf)
|
||||
.filter((sql: any) => sql.text.includes("'{retryCount}'") && sql.params.includes(imageId));
|
||||
const auditRows = () =>
|
||||
vi
|
||||
.mocked(clickhouse!.insert)
|
||||
.mock.calls.filter((call: any) => call[0].table === 'scanner_label_results')
|
||||
.flatMap((call: any) => call[0].values);
|
||||
|
||||
it('CONTROL: a POI name that is not whitelisted is still tagged', async () => {
|
||||
seedImageWithPrompt(30, 'emma stone');
|
||||
|
||||
const req = runWebhook({ id: 30, status: 0, source: TagSource.WD14, tags: [] });
|
||||
await req.promise;
|
||||
|
||||
// Deliberately asserts only the tag, with the strip passing the text through: it
|
||||
// stays green with the fix reverted, so a failure below is the strip and not a
|
||||
// broken fixture.
|
||||
expect(requestedTagNames()).toContain('emma stone');
|
||||
beforeEach(() => {
|
||||
// An earlier test leaves $executeRaw throwing, which would skip the verdict write.
|
||||
mockDbWrite.$executeRaw.mockResolvedValue(0);
|
||||
vi.mocked(clickhouse!.insert).mockClear();
|
||||
});
|
||||
|
||||
it('a whitelisted phrase is stripped before the POI check, so no POI tag is written', async () => {
|
||||
mockStripBenignPhrases.mockImplementation(async (text?: string) =>
|
||||
(text ?? '').replace('emma stone', '')
|
||||
);
|
||||
// Two real POI names, only one whitelisted. `tom hanks` survives the strip and must
|
||||
// still be tagged, which proves this run reached the tag write at all — otherwise an
|
||||
// early throw would satisfy the negative assertion below with an empty array.
|
||||
seedImageWithPrompt(31, 'emma stone and tom hanks');
|
||||
it('writes general tags, the rating and a Scanned verdict for an image', async () => {
|
||||
await deliver(40, [scanStep(scanOutput()), hashStep]);
|
||||
|
||||
const req = runWebhook({ id: 31, status: 0, source: TagSource.WD14, tags: [] });
|
||||
await req.promise;
|
||||
|
||||
// Pins WHICH list is consulted. Pointing the strip at ProfanityBenignWord instead
|
||||
// leaves the whole fix inert in production and every other assertion here green.
|
||||
expect(mockStripBenignPhrases).toHaveBeenCalledWith(
|
||||
'emma stone and tom hanks',
|
||||
'PromptBenignPhrase'
|
||||
const written = tagsWritten(40);
|
||||
expect(written).toContainEqual(
|
||||
expect.objectContaining({ tagId: 200, source: TagSource.WD14, confidence: 90 })
|
||||
);
|
||||
expect(requestedTagNames()).not.toContain('emma stone');
|
||||
expect(requestedTagNames()).toContain('tom hanks');
|
||||
expect(written).toContainEqual(
|
||||
expect.objectContaining({ tagId: 1004, source: TagSource.SpineRating })
|
||||
);
|
||||
// Other tagger categories are not ingested, even when the tag exists.
|
||||
expect(written.some((tag: any) => tag.tagId === 100 || tag.tagId === 400)).toBe(false);
|
||||
|
||||
expect(verdictFor(40)).toMatchObject({
|
||||
ingestion: ImageIngestionStatus.Scanned,
|
||||
nsfwLevel: 8,
|
||||
pHash: BigInt('0x6F51B11C49611E0E'),
|
||||
});
|
||||
expect(errorFlipsFor(40)).toHaveLength(0);
|
||||
expect(vi.mocked(signalClient.send)).toHaveBeenCalledWith(
|
||||
expect.objectContaining({
|
||||
data: expect.objectContaining({ imageId: 40, ingestion: ImageIngestionStatus.Scanned }),
|
||||
})
|
||||
);
|
||||
});
|
||||
|
||||
it('queues an NSFW image with a minor tag for review', async () => {
|
||||
await deliver(41, [scanStep(scanOutput())]);
|
||||
expect(verdictFor(41)).toMatchObject({ needsReview: 'minor', minor: true });
|
||||
});
|
||||
|
||||
it('leaves a moderator-locked NSFW level alone', async () => {
|
||||
await deliver(7, [scanStep(scanOutput())]);
|
||||
expect(verdictFor(7)).toMatchObject({
|
||||
ingestion: ImageIngestionStatus.Scanned,
|
||||
nsfwLevel: 1,
|
||||
});
|
||||
});
|
||||
|
||||
it('writes version 2 scanner-audit rows, csam included', async () => {
|
||||
await deliver(42, [scanStep(scanOutput({ csam: true })), hashStep]);
|
||||
|
||||
const rows = auditRows();
|
||||
expect(rows.map((row: any) => [row.label, row.labelValue, row.triggered])).toEqual([
|
||||
['x', 'nsfw_level', 1],
|
||||
['csam', '', 1],
|
||||
['minor', '21-24', 0],
|
||||
['ai', 'ai_recognition', 1],
|
||||
['anime', 'anime_recognition', 1],
|
||||
]);
|
||||
expect(rows.every((row: any) => row.version === '2' && row.entityIds[0] === '42')).toBe(true);
|
||||
});
|
||||
|
||||
it('rates a video by its riskiest frame', async () => {
|
||||
await deliver(43, [
|
||||
{ $type: 'videoFrameExtraction', name: 'videoFrames', status: 'succeeded', output: {} },
|
||||
{
|
||||
$type: 'repeat',
|
||||
status: 'succeeded',
|
||||
input: { template: { $type: 'imageScanning' } },
|
||||
output: {
|
||||
steps: [
|
||||
scanStep(scanOutput({ nsfwLevel: 'pg' })),
|
||||
scanStep(scanOutput({ nsfwLevel: 'x' })),
|
||||
scanStep(scanOutput({ nsfwLevel: 'pg13' })),
|
||||
],
|
||||
},
|
||||
},
|
||||
]);
|
||||
|
||||
expect(tagsWritten(43)).toContainEqual(
|
||||
expect.objectContaining({ tagId: 1004, source: TagSource.SpineRating })
|
||||
);
|
||||
expect(verdictFor(43)).toMatchObject({
|
||||
ingestion: ImageIngestionStatus.Scanned,
|
||||
nsfwLevel: 8,
|
||||
});
|
||||
});
|
||||
|
||||
it.each([
|
||||
['a failed workflow', [scanStep(undefined, 'failed')], 'failed', 'workflow-failed'],
|
||||
[
|
||||
'a scan whose tagging did not run',
|
||||
[scanStep(scanOutput({ tagging: { status: 'skipped', ran: false, tags: [] } }))],
|
||||
'succeeded',
|
||||
'unusable-result',
|
||||
],
|
||||
])('marks %s as one Error without a verdict', async (_, steps, status, failureType) => {
|
||||
await deliver(44, steps, status);
|
||||
|
||||
const flips = errorFlipsFor(44);
|
||||
expect(flips).toHaveLength(1);
|
||||
expect(flips[0].params.join(' ')).toContain(failureType);
|
||||
expect(verdictFor(44)).toBeUndefined();
|
||||
expect(tagsWritten(44)).toHaveLength(0);
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
Vendored
-3
@@ -389,10 +389,8 @@ export const serverSchema = z
|
||||
UNAUTHENTICATED_DOWNLOAD: zc.booleanString,
|
||||
UNAUTHENTICATED_LIST_NSFW: zc.booleanString,
|
||||
LOGGING: commaDelimitedStringArray(),
|
||||
IMAGE_SCANNING_ENDPOINT: isProd ? z.string() : z.string().optional(),
|
||||
IMAGE_SCANNING_CALLBACK: z.string().optional(),
|
||||
TEXT_MODERATION_CALLBACK: z.string().optional(),
|
||||
IMAGE_SCANNING_MODEL: z.string().optional(),
|
||||
IMAGE_SCANNING_RETRY_DELAY: z.coerce.number().default(5),
|
||||
// Age-out threshold (minutes) for never-returning image scans. A scan verdict
|
||||
// arrives via the fire-and-forget /image-scan-result webhook; a fraction never
|
||||
@@ -419,7 +417,6 @@ export const serverSchema = z
|
||||
// newest-first (starving the oldest backlog) — a bad hand-tune must fail
|
||||
// loudly at boot instead.
|
||||
IMAGE_SCANNING_MAX_PER_RUN: z.coerce.number().int().positive().default(1000),
|
||||
IMAGE_SCANNER_NEW: zc.booleanString.default(false),
|
||||
DELIVERY_WORKER_ENDPOINT: z.string().optional(),
|
||||
DELIVERY_WORKER_TOKEN: z.string().optional(),
|
||||
STORAGE_RESOLVER_ENDPOINT: z.string().optional(), // URL for storage-resolver microservice
|
||||
|
||||
@@ -1,78 +0,0 @@
|
||||
import { chunk } from 'lodash-es';
|
||||
import type { NextApiRequest, NextApiResponse } from 'next';
|
||||
import * as z from 'zod';
|
||||
import { ImageScanType } from '~/server/common/enums';
|
||||
import { dbRead } from '~/server/db/client';
|
||||
import { dataProcessor } from '~/server/db/db-helpers';
|
||||
import type { IngestImageInput } from '~/server/schema/image.schema';
|
||||
import { ingestImageBulk } from '~/server/services/image.service';
|
||||
import { ModEndpoint } from '~/server/utils/endpoint-helpers';
|
||||
|
||||
const schema = z.object({
|
||||
start: z.coerce.number().optional(),
|
||||
end: z.coerce.number().optional(),
|
||||
});
|
||||
|
||||
export default ModEndpoint(async function (req: NextApiRequest, res: NextApiResponse) {
|
||||
const input = schema.parse(req.query);
|
||||
const start = Date.now();
|
||||
await dataProcessor({
|
||||
params: { batchSize: 100000, concurrency: 10, start: 0 },
|
||||
runContext: {
|
||||
on: (event: 'close', listener: () => void) => {
|
||||
// noop
|
||||
},
|
||||
},
|
||||
rangeFetcher: async (ctx) => {
|
||||
let [{ start, end }] = await dbRead.$queryRaw<{ start: number; end: number }[]>`
|
||||
WITH dates AS (
|
||||
SELECT
|
||||
MIN("createdAt") as start,
|
||||
MAX("createdAt") as end
|
||||
FROM "Image"
|
||||
)
|
||||
SELECT MIN(id) as start, MAX(id) as end
|
||||
FROM "Image" i
|
||||
JOIN dates d ON d.start = i."createdAt" OR d.end = i."createdAt";
|
||||
`;
|
||||
if (input.start) start = input.start;
|
||||
if (input.end) end = input.end;
|
||||
|
||||
return { start, end };
|
||||
},
|
||||
processor: async ({ start, end }) => {
|
||||
const consoleFetchKey = `Fetch: ${start} - ${end}`;
|
||||
console.log(consoleFetchKey);
|
||||
console.time(consoleFetchKey);
|
||||
const records = await dbRead.$queryRaw<IngestImageInput[]>`
|
||||
SELECT
|
||||
"id",
|
||||
"url",
|
||||
"type",
|
||||
"height",
|
||||
"width",
|
||||
meta->>'prompt' as prompt
|
||||
FROM "Image" i
|
||||
WHERE i.id BETWEEN ${start} AND ${end}
|
||||
AND i."pHash" IS NULL
|
||||
`;
|
||||
console.timeEnd(consoleFetchKey);
|
||||
|
||||
if (records.length === 0) return;
|
||||
|
||||
const consolePushKey = `Push: ${start} - ${end}: ${records.length}`;
|
||||
console.log(consolePushKey);
|
||||
console.time(consolePushKey);
|
||||
for (const batch of chunk(records, 1000)) {
|
||||
await ingestImageBulk({
|
||||
images: batch,
|
||||
scans: [ImageScanType.Hash],
|
||||
lowPriority: true,
|
||||
});
|
||||
}
|
||||
console.timeEnd(consolePushKey);
|
||||
},
|
||||
});
|
||||
|
||||
return res.status(200).json({ success: true, duration: Date.now() - start });
|
||||
});
|
||||
@@ -1,89 +0,0 @@
|
||||
import type { NextApiRequest, NextApiResponse } from 'next';
|
||||
import { dbWrite } from '~/server/db/client';
|
||||
import * as z from 'zod';
|
||||
import { ModEndpoint } from '~/server/utils/endpoint-helpers';
|
||||
import type { Prisma } from '@prisma/client';
|
||||
import { env } from '~/env/server';
|
||||
import { chunk } from 'lodash-es';
|
||||
import { getEdgeUrl } from '~/client-utils/edge-url';
|
||||
|
||||
const stringToNumberArraySchema = z
|
||||
.string()
|
||||
.transform((s) => s.split(',').map(Number))
|
||||
.optional();
|
||||
const importSchema = z.object({
|
||||
imageCount: z.preprocess((x) => (x ? parseInt(String(x)) : undefined), z.number()).optional(),
|
||||
imageIds: stringToNumberArraySchema,
|
||||
wait: z.preprocess((val) => val === true || val === 'true', z.boolean()).optional(),
|
||||
});
|
||||
|
||||
export default ModEndpoint(
|
||||
async function scanImages(req: NextApiRequest, res: NextApiResponse) {
|
||||
if (!env.IMAGE_SCANNING_ENDPOINT)
|
||||
return res.status(400).json({ error: 'Image scanning is not enabled' });
|
||||
|
||||
const { imageCount, imageIds, wait } = importSchema.parse(req.query);
|
||||
|
||||
const where: Prisma.Enumerable<Prisma.ImageWhereInput> = {};
|
||||
if (!!imageIds?.length) where.id = { in: imageIds };
|
||||
else if (!!imageCount) {
|
||||
where.scanRequestedAt = null;
|
||||
where.scannedAt = null;
|
||||
} else {
|
||||
return res.status(400).json({
|
||||
error: 'Must provide at least one of imageCount or imageIds',
|
||||
});
|
||||
}
|
||||
|
||||
const images = await dbWrite.image.findMany({
|
||||
where,
|
||||
take: imageCount,
|
||||
select: { url: true, id: true, width: true, name: true, mimeType: true },
|
||||
});
|
||||
|
||||
if (!wait) res.status(200).json({ images: images.length });
|
||||
|
||||
const batchSize = 100;
|
||||
const batches = chunk(images, batchSize);
|
||||
let i = 0;
|
||||
for (const batch of batches) {
|
||||
console.log(
|
||||
`Sending batch ${i} to ${Math.min(i + batchSize, images.length)} of ${images.length} images`
|
||||
);
|
||||
const queued: number[] = [];
|
||||
await Promise.all(
|
||||
batch.map(async (image) => {
|
||||
const width = Math.min(450, image.width ?? 450);
|
||||
const anim =
|
||||
image.name?.endsWith('.gif') || image.mimeType == 'image/gif' ? false : undefined;
|
||||
const gamma = anim === false ? 0.99 : undefined;
|
||||
const url = getEdgeUrl(image.url, { width, anim, gamma });
|
||||
|
||||
try {
|
||||
await fetch(env.IMAGE_SCANNING_ENDPOINT as string, {
|
||||
method: 'POST',
|
||||
headers: { 'Content-Type': 'application/json' },
|
||||
body: JSON.stringify({ url, imageId: image.id }),
|
||||
});
|
||||
queued.push(image.id);
|
||||
} catch (e: any) {
|
||||
console.error('Failed to send image for scan', e.message);
|
||||
}
|
||||
})
|
||||
);
|
||||
|
||||
if (!!queued.length) {
|
||||
await dbWrite.image.updateMany({
|
||||
where: { id: { in: queued } },
|
||||
data: { scanRequestedAt: new Date() },
|
||||
});
|
||||
}
|
||||
|
||||
i += batchSize;
|
||||
}
|
||||
console.log('Done sending images for scan!');
|
||||
|
||||
if (wait) res.status(200).json({ images: images.length });
|
||||
},
|
||||
['GET']
|
||||
);
|
||||
File diff suppressed because it is too large
Load Diff
@@ -1,34 +0,0 @@
|
||||
import type { NextApiRequest, NextApiResponse } from 'next';
|
||||
import * as z from 'zod';
|
||||
import { Prisma } from '@prisma/client';
|
||||
import { dbWrite } from '~/server/db/client';
|
||||
import type { IngestImageInput } from '~/server/schema/image.schema';
|
||||
import { ingestImageBulk } from '~/server/services/image.service';
|
||||
import { WebhookEndpoint } from '~/server/utils/endpoint-helpers';
|
||||
|
||||
const schema = z.object({
|
||||
imageIds: z.array(z.number()).min(1),
|
||||
lowPriority: z.boolean().default(true),
|
||||
});
|
||||
|
||||
export default WebhookEndpoint(async function (req: NextApiRequest, res: NextApiResponse) {
|
||||
if (req.method !== 'POST') return res.status(405).json({ error: 'Method Not Allowed' });
|
||||
|
||||
const { imageIds, lowPriority } = schema.parse(req.body);
|
||||
|
||||
const images = await dbWrite.$queryRaw<IngestImageInput[]>`
|
||||
SELECT id, url, type, width, height, meta->>'prompt' as prompt
|
||||
FROM "Image"
|
||||
WHERE id IN (${Prisma.join(imageIds)})
|
||||
`;
|
||||
|
||||
if (!images.length) {
|
||||
return res.status(404).json({ error: 'No images found for the provided IDs' });
|
||||
}
|
||||
|
||||
const foundIds = images.map((i) => i.id);
|
||||
|
||||
const success = await ingestImageBulk({ images, lowPriority });
|
||||
|
||||
res.status(200).json({ success, count: images.length, imageIds: foundIds });
|
||||
});
|
||||
@@ -138,6 +138,11 @@ export enum FLIPT_FEATURE_FLAGS {
|
||||
// OFF is the shipped default and means the pattern list is recorded but not enforced on these
|
||||
// surfaces. The link-domain half throws either way — this flag has never governed it.
|
||||
USER_CONTENT_PATTERN_ENFORCE = 'user-content-pattern-enforce',
|
||||
|
||||
// Submits image ingestion as one imageScanning step instead of wdTagging + mediaRating.
|
||||
// DEFAULT-OFF — an unknown flag or unreachable Flipt keeps the two-step path. Evaluated
|
||||
// with the imageId and no context, so ramp by percentage or boolean; a segment matches nothing.
|
||||
IMAGE_INGESTION_IMAGE_SCANNING = 'image-ingestion-image-scanning',
|
||||
}
|
||||
|
||||
// Flags exempt from caching: incident kill-switches where an operator expects a
|
||||
|
||||
@@ -3,7 +3,7 @@ import { NewOrderRankType } from '~/shared/utils/prisma/enums';
|
||||
/**
|
||||
* Built-in fallback for `poolQuotas` when Redis config hasn't seeded one.
|
||||
* Knight is the only rank that suffers from the SFW/NSFW priority misuse
|
||||
* today (see `image-scan-result.ts` routing), so we default it to a 50/50
|
||||
* today (see `addToNewOrderQueue` in `image-scan-pipeline.ts`), so we default it to a 50/50
|
||||
* split between Knight1 (SFW) and Knight2 (NSFW). Acolyte/Templar are
|
||||
* intentionally omitted to preserve their legacy sequential behavior.
|
||||
*/
|
||||
|
||||
@@ -15,12 +15,9 @@ import { decreaseDate } from '~/utils/date-helpers';
|
||||
const IMAGE_SCANNING_ERROR_DELAY = 60 * 1; // 1 hour
|
||||
const IMAGE_SCANNING_RETRY_LIMIT = 9;
|
||||
|
||||
// Hard per-image backstop for a single submit. The orchestrator submit is already
|
||||
// bounded per-attempt (createImageIngestionRequest, ~15s AbortSignal), but this also
|
||||
// covers any other external await inside ingestImage (Redis flag read, prompt lookup,
|
||||
// the scanJobs UPDATE) so one hung submit ties up a single concurrency slot rather
|
||||
// than the whole run. A fired timeout just fails that image → it stays queued and is
|
||||
// retried on a later run.
|
||||
// Hard per-image backstop. The orchestrator submit is bounded per attempt (~15s), but the
|
||||
// Flipt flag read before it and the scanJobs UPDATE after it are not, so this keeps one hung
|
||||
// image to one concurrency slot. A timed-out image stays queued for a later run.
|
||||
const INGEST_IMAGE_TIMEOUT_MS = 60 * 1000;
|
||||
|
||||
// Per-run wall-clock budget. Once exceeded we stop STARTING new submits so the run
|
||||
@@ -702,9 +699,9 @@ export const removeBlockedImages = createJob(
|
||||
// DOES NOT RETRACT — the row is still hard-deleted here, exactly as before; only the shared
|
||||
// object is left alone:
|
||||
// • the scan pipeline's three block outcomes — the orchestrator content rating
|
||||
// (`blockImageFromRating`), the prompt/text audit, and the moderation rule engine — in
|
||||
// BOTH copies of that pipeline: `image-scan-result.service` and the legacy bodies in
|
||||
// `api/webhooks/image-scan-result`. Automated, no moderator, no `ModActivity`.
|
||||
// (`blockImageFromRating`), the prompt/text audit, and the moderation rule engine
|
||||
// (`image-scan-result.service` / `image-scan-pipeline`). Automated, no moderator, no
|
||||
// `ModActivity`.
|
||||
// • the CSAM branch of `report.service` — reached from `report.create`, which is a
|
||||
// `guardedProcedureAllowUnverifiedEmail`, so it is fired by ANY reporting user's report
|
||||
// and not by a moderator reviewing one. This is the writer that would be most dangerous
|
||||
|
||||
@@ -6,8 +6,8 @@
|
||||
* a read back to a utf8 string before treating it as one. No-op when the value is
|
||||
* already a string (single-node / dev) or null/undefined.
|
||||
*
|
||||
* Same coercion as the earlier per-site fixes (redis/queues.ts, image-scanner-flag.ts,
|
||||
* metrics/base.metrics.ts). Prefer this helper for new call sites.
|
||||
* Same coercion as the earlier per-site fixes (redis/queues.ts, metrics/base.metrics.ts).
|
||||
* Prefer this helper for new call sites.
|
||||
*/
|
||||
export function decodeRedisString<T extends string | null | undefined>(value: T | Buffer): T {
|
||||
return (Buffer.isBuffer(value) ? value.toString('utf8') : value) as T;
|
||||
|
||||
@@ -19,7 +19,6 @@ import {
|
||||
getImagesByUserIdForModeration,
|
||||
getImagesForModelVersionCache,
|
||||
getMyImages,
|
||||
ingestArticleCoverImages,
|
||||
ingestImageById,
|
||||
removeImageResource,
|
||||
removeImageTechniques,
|
||||
@@ -95,10 +94,6 @@ const isOwnerOrModerator = middleware(async ({ ctx, next, input = {} }) => {
|
||||
|
||||
// TODO.cleanup - remove unused router methods
|
||||
export const imageRouter = router({
|
||||
ingestArticleImages: protectedProcedure
|
||||
.meta({ requiredScope: TokenScope.MediaWrite })
|
||||
.input(z.array(z.object({ imageId: z.number(), articleId: z.number() })))
|
||||
.mutation(({ input }) => ingestArticleCoverImages(input)),
|
||||
moderate: moderatorProcedure.input(imageModerationSchema).mutation(moderateImageHandler),
|
||||
delete: verifiedProcedure
|
||||
.meta({ requiredScope: TokenScope.MediaDelete })
|
||||
|
||||
@@ -1,51 +0,0 @@
|
||||
import { describe, expect, it } from 'vitest';
|
||||
|
||||
// Regression: `isImageScannerNewEnabled` reads a kill-switch from sysRedis and
|
||||
// compared the reply against the literals '1'/'true'/'0'/'false'. The
|
||||
// HA/Sentinel sysRedis returns a Buffer for BLOB_STRING replies, which matched
|
||||
// none of the four literals → the function fell through and DESTRUCTIVELY
|
||||
// overwrote the operator's '1' with 'false' (then returned false). The pure
|
||||
// decision is extracted into its own `image-scanner-flag` module so the
|
||||
// coercion + the seed-only-on-genuinely-unset behavior is testable without
|
||||
// loading the 7.9K-line image.service module (which drags in Prisma/env/auth
|
||||
// at import time and can't load under vitest). null return == "unset" == the
|
||||
// only case the caller is allowed to default-seed.
|
||||
|
||||
import { parseScannerFlag } from '~/server/services/image-scanner-flag';
|
||||
|
||||
describe('parseScannerFlag — sysRedis Buffer-vs-string flag', () => {
|
||||
it('Buffer("1") → true (was a destructive fall-through pre-fix)', () => {
|
||||
expect(parseScannerFlag(Buffer.from('1', 'utf8'))).toBe(true);
|
||||
});
|
||||
|
||||
it('Buffer("true") → true', () => {
|
||||
expect(parseScannerFlag(Buffer.from('true', 'utf8'))).toBe(true);
|
||||
});
|
||||
|
||||
it('Buffer("0") → false', () => {
|
||||
expect(parseScannerFlag(Buffer.from('0', 'utf8'))).toBe(false);
|
||||
});
|
||||
|
||||
it('Buffer("false") → false', () => {
|
||||
expect(parseScannerFlag(Buffer.from('false', 'utf8'))).toBe(false);
|
||||
});
|
||||
|
||||
it('string "1"/"true" → true (legacy single-pod, unchanged)', () => {
|
||||
expect(parseScannerFlag('1')).toBe(true);
|
||||
expect(parseScannerFlag('true')).toBe(true);
|
||||
});
|
||||
|
||||
it('string "0"/"false" → false (legacy single-pod, unchanged)', () => {
|
||||
expect(parseScannerFlag('0')).toBe(false);
|
||||
expect(parseScannerFlag('false')).toBe(false);
|
||||
});
|
||||
|
||||
it('null → null (genuinely unset → caller seeds the default)', () => {
|
||||
expect(parseScannerFlag(null)).toBeNull();
|
||||
});
|
||||
|
||||
it('an unrecognized value → null (treated as unset, not a destructive match)', () => {
|
||||
expect(parseScannerFlag(Buffer.from('garbage', 'utf8'))).toBeNull();
|
||||
expect(parseScannerFlag('garbage')).toBeNull();
|
||||
});
|
||||
});
|
||||
@@ -0,0 +1,435 @@
|
||||
import { beforeEach, describe, expect, it, vi } from 'vitest';
|
||||
import {
|
||||
classifyImageScanFailure,
|
||||
ImageScanFailureClass,
|
||||
} from '~/server/services/image-scan-failure';
|
||||
|
||||
const pipeline = vi.hoisted(() => ({
|
||||
applyIngestionSideEffects: vi.fn(),
|
||||
buildAndInsertScanTags: vi.fn(),
|
||||
extractFailedSteps: vi.fn(),
|
||||
loadImageForScan: vi.fn(),
|
||||
logPerceptualHashMatch: vi.fn(),
|
||||
markImageScanError: vi.fn(),
|
||||
readJobFailureReason: vi.fn(),
|
||||
resolveScanOutcome: vi.fn(),
|
||||
sendIngestionSignal: vi.fn(),
|
||||
}));
|
||||
const { mockRecordAudit, mockRemoveJobQueue, mockFanOut } = vi.hoisted(() => ({
|
||||
mockRecordAudit: vi.fn(),
|
||||
mockRemoveJobQueue: vi.fn(),
|
||||
mockFanOut: vi.fn(),
|
||||
}));
|
||||
|
||||
// The pipeline imports the whole ingestion stack, and its stages run for real in
|
||||
// image-scan-result.test.ts, so this suite pins only what this service passes them.
|
||||
vi.mock('~/server/services/image-scan-pipeline', () => pipeline);
|
||||
vi.mock('~/server/services/orchestrator/orchestrator.service', () => ({
|
||||
computePerceptualHash: (hash?: string) => (hash ? BigInt(`0x${hash}`) : undefined),
|
||||
}));
|
||||
vi.mock('~/server/services/scanner-audit.service', () => ({
|
||||
recordImageScanningResult: mockRecordAudit,
|
||||
}));
|
||||
vi.mock('~/server/services/job-queue.service', () => ({
|
||||
removeImageScanJobQueue: mockRemoveJobQueue,
|
||||
}));
|
||||
vi.mock('~/server/utils/webhook-debounce', () => ({ fanOutArticleImageUpdates: mockFanOut }));
|
||||
|
||||
import {
|
||||
IMAGE_SCANNING_LOG,
|
||||
isImageScanningWorkflow,
|
||||
parseImageScanningSteps,
|
||||
processImageScanningWorkflow,
|
||||
} from '~/server/services/image-scanning-result.service';
|
||||
|
||||
const scanStep = {
|
||||
$type: 'imageScanning',
|
||||
name: 'scan',
|
||||
status: 'succeeded',
|
||||
output: {
|
||||
nsfwLevel: 'pg13',
|
||||
tagging: { ran: true, tags: [{ tag: 'solo', category: 'general', score: 0.9 }] },
|
||||
jointAgeClassification: { detections: [], minorDetected: false },
|
||||
csam: null,
|
||||
},
|
||||
};
|
||||
const failedScanStep = { ...scanStep, status: 'failed', output: undefined };
|
||||
const hashStep = { $type: 'mediaHash', output: { hashes: { perceptual: '0A' } } };
|
||||
const image = { id: 7, userId: 3, meta: { prompt: 'a prompt' } };
|
||||
|
||||
const deliver = (status: string, steps: unknown[]) =>
|
||||
processImageScanningWorkflow({
|
||||
workflowId: 'wf',
|
||||
status,
|
||||
steps,
|
||||
imageId: 7,
|
||||
articleImageScanning: true,
|
||||
startedAt: 'start',
|
||||
completedAt: 'end',
|
||||
});
|
||||
|
||||
// Every failure must be recorded exactly once: a second write bumps retryCount again and
|
||||
// overwrites the failure type the retry ceiling is chosen from.
|
||||
const expectOneScanError = (failure: Record<string, unknown>) => {
|
||||
expect(pipeline.markImageScanError).toHaveBeenCalledTimes(1);
|
||||
expect(pipeline.markImageScanError).toHaveBeenCalledWith(expect.objectContaining(failure));
|
||||
expect(pipeline.sendIngestionSignal).toHaveBeenCalledTimes(1);
|
||||
expect(pipeline.sendIngestionSignal).toHaveBeenCalledWith({
|
||||
imageId: 7,
|
||||
userId: 3,
|
||||
ingestion: 'Error',
|
||||
log: IMAGE_SCANNING_LOG,
|
||||
});
|
||||
expect(mockFanOut).toHaveBeenCalledWith(7);
|
||||
expect(pipeline.buildAndInsertScanTags).not.toHaveBeenCalled();
|
||||
};
|
||||
|
||||
describe('processImageScanningWorkflow', () => {
|
||||
beforeEach(() => {
|
||||
vi.clearAllMocks();
|
||||
pipeline.loadImageForScan.mockResolvedValue(image);
|
||||
pipeline.resolveScanOutcome.mockResolvedValue({
|
||||
ingestion: 'Scanned',
|
||||
blockedFor: null,
|
||||
reviewKey: null,
|
||||
});
|
||||
pipeline.markImageScanError.mockResolvedValue({ userId: 3, failureClass: 'unknown' });
|
||||
pipeline.extractFailedSteps.mockReturnValue(['scan']);
|
||||
pipeline.readJobFailureReason.mockResolvedValue('Failed to create container');
|
||||
});
|
||||
|
||||
it('writes the scan through the shared pipeline without a hard block', async () => {
|
||||
await deliver('succeeded', [scanStep, hashStep]);
|
||||
|
||||
expect(pipeline.logPerceptualHashMatch).toHaveBeenCalledWith({
|
||||
imageId: 7,
|
||||
pHash: 10n,
|
||||
log: IMAGE_SCANNING_LOG,
|
||||
});
|
||||
expect(pipeline.buildAndInsertScanTags).toHaveBeenCalledWith({
|
||||
imageId: 7,
|
||||
wdTags: { solo: 0.9 },
|
||||
ratingLevel: 'pg13',
|
||||
prompt: 'a prompt',
|
||||
});
|
||||
expect(pipeline.resolveScanOutcome).toHaveBeenCalledWith({
|
||||
image,
|
||||
pHash: 10n,
|
||||
workflowId: 'wf',
|
||||
prompt: 'a prompt',
|
||||
negativePrompt: undefined,
|
||||
log: IMAGE_SCANNING_LOG,
|
||||
});
|
||||
expect(mockRemoveJobQueue).toHaveBeenCalledWith([7]);
|
||||
expect(mockRecordAudit).toHaveBeenCalledWith(
|
||||
expect.objectContaining({
|
||||
workflowId: 'wf',
|
||||
imageId: 7,
|
||||
startedAt: 'start',
|
||||
completedAt: 'end',
|
||||
scan: expect.objectContaining({ nsfwLevel: 'pg13', csam: null }),
|
||||
})
|
||||
);
|
||||
expect(pipeline.applyIngestionSideEffects).toHaveBeenCalledWith({
|
||||
image,
|
||||
outcome: expect.objectContaining({ ingestion: 'Scanned' }),
|
||||
});
|
||||
expect(mockFanOut).toHaveBeenCalledWith(7);
|
||||
expect(pipeline.markImageScanError).not.toHaveBeenCalled();
|
||||
expect(pipeline.sendIngestionSignal).toHaveBeenCalledTimes(1);
|
||||
expect(pipeline.sendIngestionSignal).toHaveBeenCalledWith({
|
||||
imageId: 7,
|
||||
userId: 3,
|
||||
ingestion: 'Scanned',
|
||||
blockedFor: null,
|
||||
log: IMAGE_SCANNING_LOG,
|
||||
});
|
||||
});
|
||||
|
||||
it('signals the outcome the pipeline resolved, blocked included', async () => {
|
||||
pipeline.resolveScanOutcome.mockResolvedValue({
|
||||
ingestion: 'Blocked',
|
||||
blockedFor: 'moderated',
|
||||
reviewKey: null,
|
||||
});
|
||||
await deliver('succeeded', [scanStep]);
|
||||
|
||||
expect(pipeline.sendIngestionSignal).toHaveBeenCalledWith({
|
||||
imageId: 7,
|
||||
userId: 3,
|
||||
ingestion: 'Blocked',
|
||||
blockedFor: 'moderated',
|
||||
log: IMAGE_SCANNING_LOG,
|
||||
});
|
||||
});
|
||||
|
||||
it.each([
|
||||
['failed', 'workflow-failed'],
|
||||
['expired', 'expired'],
|
||||
['canceled', 'canceled'],
|
||||
])('marks a %s workflow as one %s error', async (status, failureType) => {
|
||||
const steps = [failedScanStep];
|
||||
await deliver(status, steps);
|
||||
|
||||
expect(pipeline.extractFailedSteps).toHaveBeenCalledWith(steps);
|
||||
expect(pipeline.readJobFailureReason).toHaveBeenCalledWith('wf');
|
||||
expect(pipeline.markImageScanError).toHaveBeenCalledWith({
|
||||
workflowId: 'wf',
|
||||
imageId: 7,
|
||||
status,
|
||||
failureType,
|
||||
failedSteps: ['scan'],
|
||||
reason: 'Failed to create container',
|
||||
});
|
||||
expectOneScanError({ failureType });
|
||||
});
|
||||
|
||||
it('marks a succeeded workflow without tags as one unusable-result error', async () => {
|
||||
await deliver('succeeded', [
|
||||
{ ...scanStep, output: { ...scanStep.output, tagging: { status: 'skipped', ran: false } } },
|
||||
]);
|
||||
|
||||
expectOneScanError({
|
||||
failureType: 'unusable-result',
|
||||
reason: expect.stringContaining('Tagging did not run'),
|
||||
});
|
||||
expect(pipeline.loadImageForScan).not.toHaveBeenCalled();
|
||||
});
|
||||
|
||||
it('marks a processing failure as one processing-failed error', async () => {
|
||||
pipeline.buildAndInsertScanTags.mockRejectedValueOnce(new Error('tag insert failed'));
|
||||
await deliver('succeeded', [scanStep]);
|
||||
|
||||
expect(pipeline.markImageScanError).toHaveBeenCalledTimes(1);
|
||||
expect(pipeline.markImageScanError).toHaveBeenCalledWith(
|
||||
expect.objectContaining({ failureType: 'processing-failed', reason: 'tag insert failed' })
|
||||
);
|
||||
expect(pipeline.sendIngestionSignal).toHaveBeenCalledTimes(1);
|
||||
expect(mockRecordAudit).not.toHaveBeenCalled();
|
||||
});
|
||||
|
||||
it.each([
|
||||
['succeeded', [scanStep]],
|
||||
['failed', [failedScanStep]],
|
||||
])('leaves articles alone when article scanning is off (%s)', async (status, steps) => {
|
||||
await processImageScanningWorkflow({
|
||||
workflowId: 'wf',
|
||||
status,
|
||||
steps,
|
||||
imageId: 7,
|
||||
articleImageScanning: false,
|
||||
});
|
||||
expect(mockFanOut).not.toHaveBeenCalled();
|
||||
});
|
||||
|
||||
it('lets a deleted image through for the route to acknowledge', async () => {
|
||||
pipeline.loadImageForScan.mockRejectedValueOnce(new Error('image not found: 7'));
|
||||
await expect(deliver('succeeded', [scanStep])).rejects.toThrow('image not found');
|
||||
expect(pipeline.markImageScanError).not.toHaveBeenCalled();
|
||||
});
|
||||
|
||||
it('keeps the verdict when a side effect fails', async () => {
|
||||
mockRemoveJobQueue.mockRejectedValueOnce(new Error('queue down'));
|
||||
await deliver('succeeded', [scanStep]);
|
||||
|
||||
expect(pipeline.markImageScanError).not.toHaveBeenCalled();
|
||||
expect(pipeline.sendIngestionSignal).toHaveBeenCalledWith(
|
||||
expect.objectContaining({ ingestion: 'Scanned' })
|
||||
);
|
||||
});
|
||||
});
|
||||
|
||||
const detection = (ageBand: string, isMinor: boolean) => ({
|
||||
ageBand,
|
||||
isMinor,
|
||||
under18Probability: isMinor ? 0.8 : 0.1,
|
||||
});
|
||||
|
||||
const scanOutput = (overrides: Record<string, unknown> = {}) => ({
|
||||
nsfwLevel: 'r',
|
||||
score: 0.8,
|
||||
topK: [],
|
||||
aiRecognition: { label: 'AI', score: 0.9 },
|
||||
animeRecognition: { label: 'real', score: 0.7 },
|
||||
tagging: {
|
||||
status: 'ran',
|
||||
ran: true,
|
||||
tags: [
|
||||
{ tag: 'solo', category: 'general', score: 0.9 },
|
||||
{ tag: 'lowres', category: 'meta', score: 0.8 },
|
||||
{ tag: 'worst quality', category: 'quality', score: 0.8 },
|
||||
{ tag: 'original', category: 'copyright', score: 0.8 },
|
||||
{ tag: 'hatsune miku', category: 'character', score: 0.8 },
|
||||
{ tag: 'questionable', category: 'rating', score: 0.8 },
|
||||
],
|
||||
},
|
||||
jointAgeClassification: {
|
||||
status: 'ran',
|
||||
ran: true,
|
||||
detections: [detection('21-24', false)],
|
||||
minorDetected: false,
|
||||
},
|
||||
csam: false,
|
||||
...overrides,
|
||||
});
|
||||
|
||||
const parserHashStep = {
|
||||
$type: 'mediaHash',
|
||||
output: { hashes: { perceptual: '6F51B11C49611E0E' } },
|
||||
};
|
||||
const parserScanStep = (output: unknown) => ({ $type: 'imageScanning', output });
|
||||
const frames = (...outputs: unknown[]) => ({
|
||||
$type: 'repeat',
|
||||
output: { steps: outputs.map(parserScanStep) },
|
||||
});
|
||||
|
||||
const untaggedFrame = scanOutput({ tagging: { status: 'skipped', ran: false, tags: [] } });
|
||||
const unratedFrame = scanOutput({ nsfwLevel: 'na' });
|
||||
|
||||
describe('parseImageScanningSteps', () => {
|
||||
it('reads an image scan, keeping only general tags', () => {
|
||||
expect(parseImageScanningSteps([parserScanStep(scanOutput()), parserHashStep], 'wf')).toEqual({
|
||||
nsfwLevel: 'r',
|
||||
tags: { solo: 0.9 },
|
||||
csam: false,
|
||||
ageDetections: [detection('21-24', false)],
|
||||
minorDetected: false,
|
||||
aiRecognition: { label: 'AI', score: 0.9 },
|
||||
animeRecognition: { label: 'real', score: 0.7 },
|
||||
perceptualHash: '6F51B11C49611E0E',
|
||||
});
|
||||
});
|
||||
|
||||
const riskyFrame = scanOutput({
|
||||
nsfwLevel: 'x',
|
||||
csam: true,
|
||||
aiRecognition: { label: 'AI', score: 0.9 },
|
||||
animeRecognition: { label: 'anime', score: 0.95 },
|
||||
tagging: { ran: true, tags: [{ tag: 'solo', category: 'general', score: 0.7 }] },
|
||||
jointAgeClassification: {
|
||||
ran: true,
|
||||
detections: [detection('13-15', true)],
|
||||
minorDetected: true,
|
||||
},
|
||||
});
|
||||
const safeFrame = scanOutput({
|
||||
nsfwLevel: 'pg',
|
||||
csam: null,
|
||||
aiRecognition: { label: 'Real', score: 0.6 },
|
||||
animeRecognition: { label: 'real', score: 0.55 },
|
||||
tagging: { ran: true, tags: [{ tag: 'solo', category: 'general', score: 0.6 }] },
|
||||
jointAgeClassification: {
|
||||
ran: true,
|
||||
detections: [detection('21-24', false)],
|
||||
minorDetected: false,
|
||||
},
|
||||
});
|
||||
|
||||
// Both orders, so a frame can only ever raise what the video reports.
|
||||
it.each([
|
||||
['riskiest frame first', [riskyFrame, safeFrame]],
|
||||
['riskiest frame last', [safeFrame, riskyFrame]],
|
||||
])('combines video frames by their highest values (%s)', (_, frameOutputs) => {
|
||||
const scan = parseImageScanningSteps(
|
||||
[{ $type: 'videoFrameExtraction', output: { frames: [] } }, frames(...frameOutputs)],
|
||||
'wf'
|
||||
);
|
||||
|
||||
expect(scan).toMatchObject({
|
||||
nsfwLevel: 'x',
|
||||
tags: { solo: 0.7 },
|
||||
csam: true,
|
||||
minorDetected: true,
|
||||
aiRecognition: { label: 'AI', score: 0.9 },
|
||||
animeRecognition: { label: 'anime', score: 0.95 },
|
||||
});
|
||||
expect(scan.ageDetections).toHaveLength(2);
|
||||
expect(scan.ageDetections).toEqual(
|
||||
expect.arrayContaining([detection('13-15', true), detection('21-24', false)])
|
||||
);
|
||||
});
|
||||
|
||||
it('reports csam as absent only when no frame reported it', () => {
|
||||
const csamOf = (...values: unknown[]) =>
|
||||
parseImageScanningSteps([frames(...values.map((csam) => scanOutput({ csam })))], 'wf').csam;
|
||||
expect(csamOf(null, null)).toBeNull();
|
||||
expect(csamOf(null, false)).toBe(false);
|
||||
expect(csamOf(false, true, null)).toBe(true);
|
||||
expect(csamOf(true, false)).toBe(true);
|
||||
});
|
||||
|
||||
it.each([
|
||||
['tagging did not run', [parserScanStep(untaggedFrame)], 'Tagging did not run'],
|
||||
['the rating is unavailable', [parserScanStep(unratedFrame)], 'media rating unavailable'],
|
||||
[
|
||||
'the rating is one this app does not know',
|
||||
[parserScanStep(scanOutput({ nsfwLevel: 'nc17' }))],
|
||||
'media rating unavailable (nc17)',
|
||||
],
|
||||
['one frame was not tagged', [frames(scanOutput(), untaggedFrame)], 'Tagging did not run'],
|
||||
['one frame has no rating', [frames(scanOutput(), unratedFrame)], 'media rating unavailable'],
|
||||
[
|
||||
'the scan step has no output',
|
||||
[{ $type: 'imageScanning', status: 'failed' }],
|
||||
'Missing imageScanning output',
|
||||
],
|
||||
[
|
||||
'a frame has no output',
|
||||
[
|
||||
{
|
||||
$type: 'repeat',
|
||||
output: { steps: [parserScanStep(scanOutput()), { $type: 'imageScanning' }] },
|
||||
},
|
||||
],
|
||||
'Missing imageScanning output',
|
||||
],
|
||||
['no frames were extracted', [frames()], 'Missing imageScanning output'],
|
||||
])('throws when %s', (_, steps, message) => {
|
||||
expect(() => parseImageScanningSteps(steps, 'wf')).toThrow(message);
|
||||
});
|
||||
|
||||
// A permanent class stops retries after one attempt; none of these is a bad file.
|
||||
it.each([[[parserScanStep(untaggedFrame)]], [[parserScanStep(unratedFrame)]], [[frames()]]])(
|
||||
'throws an error the retry classifier keeps retrying',
|
||||
(steps) => {
|
||||
let reason = '';
|
||||
try {
|
||||
parseImageScanningSteps(steps, 'wf');
|
||||
} catch (error) {
|
||||
reason = (error as Error).message;
|
||||
}
|
||||
expect(reason).not.toBe('');
|
||||
expect(
|
||||
classifyImageScanFailure({ reason, failureType: 'unusable-result', failedSteps: [] })
|
||||
).not.toBe(ImageScanFailureClass.Permanent);
|
||||
}
|
||||
);
|
||||
});
|
||||
|
||||
describe('isImageScanningWorkflow', () => {
|
||||
it.each([
|
||||
['a finished image scan', [parserScanStep(scanOutput()), parserHashStep], true],
|
||||
['a failed image scan with no output', [{ $type: 'imageScanning', status: 'failed' }], true],
|
||||
['a finished video scan', [frames(scanOutput())], true],
|
||||
[
|
||||
'a video scan that never produced frames',
|
||||
[{ $type: 'repeat', input: { template: { $type: 'imageScanning' } } }],
|
||||
true,
|
||||
],
|
||||
[
|
||||
'a legacy image scan',
|
||||
[{ $type: 'wdTagging' }, { $type: 'mediaRating' }, parserHashStep],
|
||||
false,
|
||||
],
|
||||
[
|
||||
'a legacy video scan',
|
||||
[
|
||||
{ $type: 'repeat', input: { template: { $type: 'wdTagging' } } },
|
||||
{ $type: 'repeat', input: { template: { $type: 'mediaRating' } } },
|
||||
],
|
||||
false,
|
||||
],
|
||||
])('%s → %s', (_, steps, expected) => {
|
||||
expect(isImageScanningWorkflow(steps)).toBe(expected);
|
||||
});
|
||||
});
|
||||
@@ -7,7 +7,7 @@ import path from 'path';
|
||||
* must read the benign-stripped prompt.
|
||||
*
|
||||
* This exists because the fix for that bug was applied to one of two live copies. The image
|
||||
* scan pipeline has two implementations — `getTagsFromIncomingTags` in the webhook (legacy
|
||||
* scan pipeline had two implementations —`getTagsFromIncomingTags` in the webhook (legacy
|
||||
* `TagSource.WD14`/`Clavata`/`Hive` bodies) and `processTags` in the service (bodies carrying
|
||||
* `workflowId`, i.e. the orchestrator) — and the webhook dispatches between them by body
|
||||
* shape. Patching the one the audit named left the other writing a confidence-100 POI tag
|
||||
@@ -33,8 +33,8 @@ import path from 'path';
|
||||
* a new one written as an arrow needs the strip INLINE at the call to be checked properly.
|
||||
* The escape also weakens as the enclosing function grows, since ANY `stripBenignPhrases`
|
||||
* between the declaration and the call satisfies it. An entry whose call strips inline
|
||||
* should therefore set `requireInline` and give up the escape entirely — only the two scan
|
||||
* files need it, because they shadow `prompt` with the stripped copy further up.
|
||||
* should therefore set `requireInline` and give up the escape entirely — only the scan
|
||||
* pipeline needs it, because it shadows `prompt` with the stripped copy further up.
|
||||
* 2. One other server-side `includesPoi` call reads a raw prompt and is deliberately NOT
|
||||
* covered: `services/apps/shared-content-safety.ts` throws `SharedContentBlockedError('poi')`
|
||||
* on shared app content, which is a hard legal reject rather than a moderation label.
|
||||
@@ -53,13 +53,7 @@ const REPO_ROOT = path.resolve(__dirname, '../../../..');
|
||||
*/
|
||||
const GUARDED_FILES: { rel: string; consequence: string; requireInline?: boolean }[] = [
|
||||
{
|
||||
rel: 'src/pages/api/webhooks/image-scan-result.ts',
|
||||
consequence:
|
||||
'writes a confidence-100 POI tag that reaches the image search index, so a ' +
|
||||
'moderator-whitelisted phrase would still flag the image',
|
||||
},
|
||||
{
|
||||
rel: 'src/server/services/image-scan-result.service.ts',
|
||||
rel: 'src/server/services/image-scan-pipeline.ts',
|
||||
consequence:
|
||||
'writes a confidence-100 POI tag that reaches the image search index, so a ' +
|
||||
'moderator-whitelisted phrase would still flag the image',
|
||||
@@ -68,7 +62,7 @@ const GUARDED_FILES: { rel: string; consequence: string; requireInline?: boolean
|
||||
rel: 'src/server/services/orchestrator/orchestration-new.service.ts',
|
||||
// The enclosing function is ~300 lines, so the scope escape below would be satisfied by ANY
|
||||
// later `stripBenignPhrases` in it — negative-prompt stripping being the obvious one, since
|
||||
// both scan files pair it with the prompt strip. This call needs no escape, so it gets none.
|
||||
// the scan pipeline pairs it with the prompt strip. This call needs no escape, so it gets none.
|
||||
requireInline: true,
|
||||
consequence:
|
||||
'refuses the generation outright for a viewer with `disablePoi`, so a ' +
|
||||
|
||||
@@ -0,0 +1,88 @@
|
||||
import { beforeEach, describe, expect, it, vi } from 'vitest';
|
||||
import type * as ClickhouseClient from '~/server/clickhouse/client';
|
||||
|
||||
const { mockInsert } = vi.hoisted(() => ({ mockInsert: vi.fn() }));
|
||||
|
||||
vi.mock('~/server/clickhouse/client', async (importOriginal) => ({
|
||||
...(await importOriginal<typeof ClickhouseClient>()),
|
||||
clickhouse: { insert: mockInsert },
|
||||
}));
|
||||
|
||||
import { recordImageScanningResult } from '~/server/services/scanner-audit.service';
|
||||
|
||||
const detection = (ageBand: string, under18Probability: number, isMinor: boolean) =>
|
||||
({ ageBand, under18Probability, isMinor } as never);
|
||||
|
||||
const scan = {
|
||||
nsfwLevel: 'r' as const,
|
||||
tags: {},
|
||||
csam: false,
|
||||
// The highest under-18 probability is not the first detection.
|
||||
ageDetections: [detection('21-24', 0.1, false), detection('16-17', 0.7, true)],
|
||||
minorDetected: true,
|
||||
aiRecognition: { label: 'AI', score: 0.9 },
|
||||
animeRecognition: { label: 'Real', score: 0.6 },
|
||||
};
|
||||
|
||||
type Row = Record<string, unknown>;
|
||||
const insertedRows = () => mockInsert.mock.calls[0][0].values as Row[];
|
||||
const row = (label: string) => insertedRows().find((r) => r.label === label);
|
||||
|
||||
describe('recordImageScanningResult', () => {
|
||||
beforeEach(() => mockInsert.mockReset().mockResolvedValue(undefined));
|
||||
|
||||
it('writes version 2 rows for the rating, csam, the joint age model and both recognizers', async () => {
|
||||
await recordImageScanningResult({ workflowId: 'wf', imageId: 1, scan });
|
||||
|
||||
const rows = insertedRows();
|
||||
expect(rows.map((r) => [r.label, r.labelValue, r.score, r.triggered])).toEqual([
|
||||
['r', 'nsfw_level', 1, 1],
|
||||
['csam', '', 0, 0],
|
||||
['minor', '21-24, 16-17', 0.7, 1],
|
||||
['ai', 'ai_recognition', 0.9, 1],
|
||||
['real', 'anime_recognition', 0.6, 1],
|
||||
]);
|
||||
expect(rows.every((r) => r.version === '2' && r.modelVersion === '2')).toBe(true);
|
||||
expect(rows.every((r) => r.scanner === 'image_ingestion')).toBe(true);
|
||||
});
|
||||
|
||||
it('records a positive csam result as triggered', async () => {
|
||||
await recordImageScanningResult({
|
||||
workflowId: 'wf',
|
||||
imageId: 1,
|
||||
scan: { ...scan, csam: true },
|
||||
});
|
||||
expect(row('csam')).toMatchObject({ score: 1, triggered: 1 });
|
||||
});
|
||||
|
||||
it('writes no csam row when the scanner did not report one', async () => {
|
||||
await recordImageScanningResult({
|
||||
workflowId: 'wf',
|
||||
imageId: 1,
|
||||
scan: { ...scan, csam: null },
|
||||
});
|
||||
expect(row('csam')).toBeUndefined();
|
||||
});
|
||||
|
||||
it('does not trigger the minor row when only adults were detected', async () => {
|
||||
await recordImageScanningResult({
|
||||
workflowId: 'wf',
|
||||
imageId: 1,
|
||||
scan: {
|
||||
...scan,
|
||||
ageDetections: [detection('21-24', 0.1, false), detection('25+', 0.2, false)],
|
||||
minorDetected: false,
|
||||
},
|
||||
});
|
||||
expect(row('minor')).toMatchObject({ labelValue: '21-24, 25+', score: 0.2, triggered: 0 });
|
||||
});
|
||||
|
||||
it('writes no minor row when nobody was detected', async () => {
|
||||
await recordImageScanningResult({
|
||||
workflowId: 'wf',
|
||||
imageId: 1,
|
||||
scan: { ...scan, ageDetections: [], minorDetected: false },
|
||||
});
|
||||
expect(row('minor')).toBeUndefined();
|
||||
});
|
||||
});
|
||||
@@ -1863,7 +1863,7 @@ export type ArticleTextModerationStatus = {
|
||||
updatedAt: Date | null;
|
||||
};
|
||||
|
||||
// `scanJobs.error` is stamped by `markImageScanError` (image-scan-result.service, scan
|
||||
// `scanJobs.error` is stamped by `markImageScanError` (image-scan-pipeline, scan
|
||||
// verdicts) and `markImageScanSubmitFailure` (image.service, submit rejections). Both
|
||||
// carry the classifier's verdict (transient | permanent | unknown) plus the human reason,
|
||||
// letting the scan-status UI render a class-aware cause.
|
||||
|
||||
@@ -1637,7 +1637,7 @@ export async function getImagesQueue({
|
||||
|
||||
// Per-rank pool weights drive a stratified fetch instead of the legacy
|
||||
// strict-sequential drain. Knight1/Knight2/Knight3 are content-tier
|
||||
// buckets in practice (see image-scan-result.ts), so reading Knight1
|
||||
// buckets in practice (see addToNewOrderQueue in image-scan-pipeline.ts), so reading Knight1
|
||||
// until full starved Knight2 (NSFW) entirely. Resolve weights from
|
||||
// Redis (ops-tunable) with a built-in fallback; missing rank → legacy.
|
||||
const rateLimitConfig = await getVotingRateLimitConfig();
|
||||
|
||||
@@ -0,0 +1,962 @@
|
||||
import poiWords from '~/utils/metadata/lists/words-poi.json';
|
||||
import { dbWrite } from '~/server/db/client';
|
||||
import { clickhouse } from '~/server/clickhouse/client';
|
||||
import { env } from '~/env/server';
|
||||
import type { TagType } from '~/shared/utils/prisma/enums';
|
||||
import {
|
||||
ImageIngestionStatus,
|
||||
ModerationRuleAction,
|
||||
NewOrderRankType,
|
||||
TagSource,
|
||||
} from '~/shared/utils/prisma/enums';
|
||||
import {
|
||||
BlockedReason,
|
||||
BlocklistType,
|
||||
NotificationCategory,
|
||||
NsfwLevel,
|
||||
SearchIndexUpdateQueueAction,
|
||||
SignalMessages,
|
||||
} from '~/server/common/enums';
|
||||
import { stripBenignPhrases } from '~/server/services/blocklist.service';
|
||||
import {
|
||||
auditMetaData,
|
||||
getTagsFromPrompt,
|
||||
includesInappropriate,
|
||||
includesPoi,
|
||||
} from '~/utils/metadata/audit';
|
||||
import { getComputedTags, getConditionalTagsForReview } from '~/server/utils/tag-rules';
|
||||
import { getTagRules } from '~/server/services/system-cache';
|
||||
import { Prisma } from '@prisma/client';
|
||||
import { insertTagsOnImageNew } from '~/server/services/tagsOnImageNew.service';
|
||||
import { isDefined } from '~/utils/type-guards';
|
||||
import { normalizeText } from '~/utils/normalize-text';
|
||||
import { styleTags, tagsNeedingReview } from '~/libs/tags';
|
||||
import { sfwBrowsingLevelsFlag } from '~/shared/constants/browsingLevel.constants';
|
||||
import { createImageTagsForReview } from '~/server/services/image-review.service';
|
||||
import {
|
||||
tagIdsForImagesCache,
|
||||
tagCacheByName,
|
||||
userImageVideoCountCaches,
|
||||
} from '~/server/redis/caches';
|
||||
import type { RedisKeyTemplateSys } from '~/server/redis/client';
|
||||
import { REDIS_SYS_KEYS, sysRedis } from '~/server/redis/client';
|
||||
import { classifyImageScanFailure } from '~/server/services/image-scan-failure';
|
||||
import type { MediaMetadata } from '~/server/schema/media.schema';
|
||||
import { deleteUserProfilePictureCache } from '~/server/services/user.service';
|
||||
import { bustCachesForPosts, updatePostNsfwLevel } from '~/server/services/post.service';
|
||||
import {
|
||||
queueComicsForPanelImage,
|
||||
updateComicNsfwLevelsForImage,
|
||||
updateModel3DNsfwLevelForThumbnailImage,
|
||||
} from '~/server/services/nsfwLevels.service';
|
||||
import { getImagesModRules, queueImageSearchIndexUpdate } from '~/server/services/image.service';
|
||||
import { signalClient } from '~/utils/signal-client';
|
||||
import { addImageToQueue } from '~/server/services/games/new-order.service';
|
||||
import { logToAxiom } from '~/server/logging/client';
|
||||
import { evaluateRules } from '~/server/utils/mod-rules';
|
||||
import { createNotification } from '~/server/services/notification.service';
|
||||
import { decreaseDate } from '~/utils/date-helpers';
|
||||
import { withRetries } from '~/utils/errorHandling';
|
||||
|
||||
// Ingestion stages shared by the legacy (wdTagging + mediaRating) and imageScanning result handlers.
|
||||
|
||||
type NormalizedTag = {
|
||||
name: string;
|
||||
confidence: number;
|
||||
source: TagSource;
|
||||
};
|
||||
|
||||
type TagWithId = { id: number; name: string; nsfwLevel: number; type: TagType };
|
||||
type ProcessedTag = {
|
||||
source: TagSource;
|
||||
confidence: number;
|
||||
id: number;
|
||||
name: string;
|
||||
nsfwLevel: number;
|
||||
type: TagType;
|
||||
};
|
||||
|
||||
// TTL for a stashed job reason. Comfortably longer than the orchestrator's 10-min
|
||||
// workflow expiry so the terminal workflow event can still read it.
|
||||
const JOB_REASON_TTL_SECONDS = 15 * 60;
|
||||
|
||||
const jobReasonKey = (workflowId: string) =>
|
||||
`${REDIS_SYS_KEYS.WEBHOOKS.IMAGE_SCAN_JOB_REASON}:${workflowId}` as RedisKeyTemplateSys;
|
||||
|
||||
/** Orchestrator job-level event shape we care about (a subset of WorkflowStepJobEvent). */
|
||||
export type OrchestratorJobEvent = {
|
||||
$type?: string;
|
||||
workflowId?: string;
|
||||
jobId?: string;
|
||||
reason?: string | null;
|
||||
};
|
||||
|
||||
// Stash the job's failure `reason` keyed by workflowId. No DB/orchestrator round-trip:
|
||||
// just a short-lived Redis write so the terminal workflow event can classify.
|
||||
//
|
||||
// Correlation contract: job events fire before the terminal workflow event, so the
|
||||
// reason is stashed by the time the workflow event reads it. Across a workflow's
|
||||
// several jobs this is last-write-wins (fine — image scans typically have one failing
|
||||
// job; the reason strings we classify on are equivalent). If the reason is missing
|
||||
// (race, or a job event that never arrived) the failure classifies as Unknown, which
|
||||
// retries conservatively under the bounded cap — safe by construction. The TTL bounds
|
||||
// a stash that never gets a following workflow event so it can't linger.
|
||||
export async function captureJobFailureReason(event: OrchestratorJobEvent) {
|
||||
const workflowId = event.workflowId;
|
||||
const reason = typeof event.reason === 'string' ? event.reason.trim() : '';
|
||||
if (!workflowId || !reason) return;
|
||||
await sysRedis
|
||||
.set(jobReasonKey(workflowId), reason, { EX: JOB_REASON_TTL_SECONDS })
|
||||
.catch(() => null);
|
||||
}
|
||||
|
||||
export async function readJobFailureReason(workflowId: string): Promise<string | null> {
|
||||
const reason = await sysRedis.get(jobReasonKey(workflowId)).catch(() => null);
|
||||
if (reason) sysRedis.del(jobReasonKey(workflowId)).catch(() => null);
|
||||
return reason ?? null;
|
||||
}
|
||||
|
||||
/** Axiom `name`/`source` a pipeline's logs carry, so each pipeline can be queried alone. */
|
||||
export type ScanLog = { name: string; source: string };
|
||||
export const LEGACY_SCAN_LOG: ScanLog = {
|
||||
name: 'image-scan-result',
|
||||
source: 'image-scan-result.service',
|
||||
};
|
||||
|
||||
/**
|
||||
* Push the resolved ingestion state to the uploader's open editor. The editor renders
|
||||
* Pending as an in-progress spinner, so any state that LEAVES Pending has to send —
|
||||
* Error included, retryable though it is — or the card goes on claiming to analyze
|
||||
* until the page is reloaded.
|
||||
*/
|
||||
export async function sendIngestionSignal({
|
||||
imageId,
|
||||
userId,
|
||||
ingestion,
|
||||
blockedFor,
|
||||
log = LEGACY_SCAN_LOG,
|
||||
}: {
|
||||
imageId: number;
|
||||
userId: number | null;
|
||||
ingestion: ImageIngestionStatus;
|
||||
blockedFor?: string | null;
|
||||
log?: ScanLog;
|
||||
}) {
|
||||
if (!userId) return;
|
||||
await signalClient
|
||||
.send({
|
||||
target: SignalMessages.ImageIngestionStatus,
|
||||
data: { imageId, ingestion, blockedFor },
|
||||
userId,
|
||||
})
|
||||
.catch((error) =>
|
||||
logToAxiom(
|
||||
{
|
||||
name: log.name,
|
||||
type: 'warning',
|
||||
message: `signal send failed: ${
|
||||
error instanceof Error ? error.message : 'Unknown error'
|
||||
}`,
|
||||
imageId,
|
||||
source: log.source,
|
||||
},
|
||||
'webhooks'
|
||||
).catch(() => null)
|
||||
);
|
||||
}
|
||||
|
||||
// Blocked-content detection
|
||||
// --------------------------------------------------
|
||||
async function getIsImageBlocked(hash: bigint) {
|
||||
if (!env.BLOCKED_IMAGE_HASH_CHECK || !clickhouse) return false;
|
||||
|
||||
const client = clickhouse;
|
||||
// $query has no retry of its own, and the observed failure is a dropped socket.
|
||||
const rows = await withRetries(
|
||||
() => client.$query<{ count: number }>`
|
||||
SELECT cast(count() as int) as count
|
||||
FROM blocked_images
|
||||
WHERE bitCount(bitXor(hash, ${hash})) < 5 AND disabled = false
|
||||
`,
|
||||
2,
|
||||
250
|
||||
);
|
||||
|
||||
return (rows?.[0]?.count ?? 0) > 0;
|
||||
}
|
||||
|
||||
export async function logPerceptualHashMatch({
|
||||
imageId,
|
||||
pHash,
|
||||
log = LEGACY_SCAN_LOG,
|
||||
}: {
|
||||
imageId: number;
|
||||
pHash: bigint;
|
||||
log?: ScanLog;
|
||||
}) {
|
||||
// Nothing branches on this result, so an exhausted retry must not fail the webhook and
|
||||
// discard a scan that already produced tags and a rating.
|
||||
const pHashBlocked = await getIsImageBlocked(pHash).catch((error) => {
|
||||
logToAxiom(
|
||||
{
|
||||
name: 'image-phash-match',
|
||||
type: 'warning',
|
||||
message: 'pHash blocklist check failed',
|
||||
imageId,
|
||||
error: error instanceof Error ? error.message : 'Unknown error',
|
||||
source: log.source,
|
||||
},
|
||||
'webhooks'
|
||||
).catch(() => null);
|
||||
return false;
|
||||
});
|
||||
if (!pHashBlocked) return;
|
||||
|
||||
// blockedReason = 'Similar to blocked content';
|
||||
logToAxiom(
|
||||
{
|
||||
name: 'image-phash-match',
|
||||
type: 'info',
|
||||
message: 'Image pHash matched a blocked image',
|
||||
imageId,
|
||||
pHash: pHash.toString(),
|
||||
source: log.source,
|
||||
},
|
||||
'webhooks'
|
||||
).catch(() => null);
|
||||
}
|
||||
|
||||
/**
|
||||
* Which orchestrator steps reported `failed`, by name (else `$type`). The workflow status
|
||||
* carries no reason, so this is the only way to tell a scan-step failure from a `hash` failure.
|
||||
*/
|
||||
export function extractFailedSteps(steps: unknown[]): string[] {
|
||||
return (steps as Array<{ name?: string; $type?: string; status?: string }>)
|
||||
.filter((step) => step?.status === 'failed')
|
||||
.map((step) => step.name ?? step.$type ?? 'unknown');
|
||||
}
|
||||
|
||||
/**
|
||||
* Flip an image to `Error`, increment its scan `retryCount`, and stamp a small
|
||||
* `scanJobs.error = { status, failureType, failedSteps, reason, failureClass, at }`
|
||||
* so a plain Postgres query can tell WHY a scan errored (and which step) without an
|
||||
* orchestrator lookup — and so the `ingest-images` cron can pick a retry ceiling
|
||||
* from `failureClass`. retryCount ALWAYS increments: it's the absolute attempt
|
||||
* count the per-class ceilings are applied against. Returns the new (post-increment)
|
||||
* retryCount, the image's mediaType, and the computed failureClass; retryCount /
|
||||
* mediaType are `null` when no row matched (e.g. the image was deleted between scan
|
||||
* request and callback).
|
||||
*/
|
||||
export async function markImageScanError({
|
||||
workflowId,
|
||||
imageId,
|
||||
status,
|
||||
failureType,
|
||||
failedSteps,
|
||||
reason,
|
||||
middleware,
|
||||
}: {
|
||||
workflowId: string;
|
||||
imageId: number;
|
||||
status: string;
|
||||
failureType: string;
|
||||
failedSteps: string[];
|
||||
/** Human failure reason from the job-level callback, if captured. */
|
||||
reason?: string | null;
|
||||
/** Orchestrator middleware locus, if known. Not exposed on the v2 job event today. */
|
||||
middleware?: string | null;
|
||||
}): Promise<{
|
||||
retryCount: number | null;
|
||||
mediaType: string | null;
|
||||
userId: number | null;
|
||||
failureClass: string;
|
||||
}> {
|
||||
const failureClass = classifyImageScanFailure({ reason, failureType, middleware, failedSteps });
|
||||
// undefined keys are dropped by JSON.stringify, so absent reason/middleware
|
||||
// simply don't appear in the stored blob.
|
||||
const errorJson = JSON.stringify({
|
||||
status,
|
||||
failureType,
|
||||
failedSteps,
|
||||
reason: reason ?? undefined,
|
||||
middleware: middleware ?? undefined,
|
||||
failureClass,
|
||||
at: new Date().toISOString(),
|
||||
});
|
||||
const rows = await dbWrite.$queryRaw<
|
||||
{ retryCount: number | null; mediaType: string | null; userId: number | null }[]
|
||||
>`
|
||||
UPDATE "Image"
|
||||
SET
|
||||
"ingestion" = ${ImageIngestionStatus.Error}::"ImageIngestionStatus",
|
||||
"scanJobs" = jsonb_set(
|
||||
jsonb_set(
|
||||
jsonb_set(
|
||||
COALESCE("scanJobs", '{}'),
|
||||
'{retryCount}',
|
||||
to_jsonb(COALESCE(("scanJobs"->>'retryCount')::int, 0) + 1)
|
||||
),
|
||||
'{workflowId}',
|
||||
${JSON.stringify(workflowId)}::jsonb
|
||||
),
|
||||
'{error}',
|
||||
${errorJson}::jsonb
|
||||
)
|
||||
WHERE id = ${imageId}
|
||||
RETURNING ("scanJobs"->>'retryCount')::int as "retryCount", type as "mediaType",
|
||||
"userId"
|
||||
`;
|
||||
return {
|
||||
retryCount: rows[0]?.retryCount ?? null,
|
||||
mediaType: rows[0]?.mediaType ?? null,
|
||||
userId: rows[0]?.userId ?? null,
|
||||
failureClass,
|
||||
};
|
||||
}
|
||||
|
||||
// Image loading
|
||||
// --------------------------------------------------
|
||||
export async function loadImageForScan(imageId: number) {
|
||||
const image = await dbWrite.image.findUnique({
|
||||
where: { id: imageId },
|
||||
select: {
|
||||
id: true,
|
||||
createdAt: true,
|
||||
scannedAt: true,
|
||||
type: true,
|
||||
userId: true,
|
||||
meta: true,
|
||||
metadata: true,
|
||||
postId: true,
|
||||
nsfwLevelLocked: true,
|
||||
nsfwLevel: true,
|
||||
ingestion: true,
|
||||
},
|
||||
});
|
||||
|
||||
if (!image) throw new Error(`image not found: ${imageId}`);
|
||||
return image;
|
||||
}
|
||||
export type ScanImage = Awaited<ReturnType<typeof loadImageForScan>>;
|
||||
|
||||
// Tagging
|
||||
// --------------------------------------------------
|
||||
export async function buildAndInsertScanTags({
|
||||
imageId,
|
||||
wdTags,
|
||||
ratingLevel,
|
||||
prompt,
|
||||
}: {
|
||||
imageId: number;
|
||||
wdTags: Record<string, number>;
|
||||
ratingLevel: string;
|
||||
prompt?: string;
|
||||
}) {
|
||||
const tagsWithSource = {
|
||||
[TagSource.WD14]: wdTags,
|
||||
[TagSource.SpineRating]: { [ratingLevel]: 100 },
|
||||
};
|
||||
const normalizedTags: NormalizedTag[] = Object.entries(tagsWithSource).flatMap(
|
||||
([source, tagMap]) =>
|
||||
Object.entries(tagMap).map(([name, confidence]) => {
|
||||
if (source === TagSource.WD14) name = name.replace(/_/g, ' ');
|
||||
return {
|
||||
name,
|
||||
confidence: Math.round(confidence * 100),
|
||||
source: source as TagSource,
|
||||
};
|
||||
})
|
||||
);
|
||||
|
||||
const tags = await processTags({ tags: normalizedTags, prompt });
|
||||
|
||||
await insertTagsOnImageNew(
|
||||
tags.map((tag) => ({
|
||||
imageId,
|
||||
tagId: tag.id,
|
||||
source: tag.source,
|
||||
confidence: tag.confidence,
|
||||
automated: true,
|
||||
}))
|
||||
);
|
||||
}
|
||||
|
||||
async function processTags({
|
||||
tags: normalized,
|
||||
prompt,
|
||||
}: {
|
||||
tags: NormalizedTag[];
|
||||
prompt?: string;
|
||||
}): Promise<ProcessedTag[]> {
|
||||
if (prompt) {
|
||||
// Strip moderator-whitelisted phrases first, or a whitelisted proper noun is written as a
|
||||
// confidence-100 POI tag that reaches the image search index.
|
||||
const realPersonName = includesPoi(
|
||||
// Normalized first, as both audit paths do. Stripping the raw text while the audit that
|
||||
// runs next reads the normalized copy means two different alphabets decide what counts
|
||||
// as whitelisted for the same prompt.
|
||||
await stripBenignPhrases(normalizeText(prompt), BlocklistType.PromptBenignPhrase)
|
||||
);
|
||||
if (realPersonName) {
|
||||
const tagName =
|
||||
typeof realPersonName === 'object' ? realPersonName.matchedText : realPersonName;
|
||||
normalized.push({
|
||||
name: tagName.toLowerCase(),
|
||||
confidence: 100,
|
||||
source: TagSource.Computed,
|
||||
});
|
||||
}
|
||||
|
||||
// Detect tags from prompt
|
||||
const promptTags = getTagsFromPrompt(prompt);
|
||||
if (promptTags)
|
||||
normalized.push(
|
||||
...promptTags.map((name) => ({ name, confidence: 70, source: TagSource.Computed }))
|
||||
);
|
||||
}
|
||||
|
||||
// add computed tags
|
||||
const computedTags = getComputedTags(
|
||||
normalized.map((x) => x.name),
|
||||
'WD14'
|
||||
);
|
||||
normalized.push(
|
||||
...computedTags.map((name) => ({ name, confidence: 70, source: TagSource.Computed }))
|
||||
);
|
||||
|
||||
// apply tag rules
|
||||
const tagRules = await getTagRules();
|
||||
for (const rule of tagRules) {
|
||||
const match = normalized.find((x) => x.name === rule.toTag);
|
||||
if (!match) continue;
|
||||
|
||||
if (rule.type === 'Replace') {
|
||||
match.name = rule.fromTag;
|
||||
} else if (rule.type === 'Append') {
|
||||
normalized.push({ name: rule.fromTag, confidence: 70, source: TagSource.Computed });
|
||||
}
|
||||
}
|
||||
|
||||
// De-dupe incoming tags and keep tag with highest confidence
|
||||
const tagMap: Record<string, NormalizedTag> = {};
|
||||
for (const tag of normalized) {
|
||||
if (!tagMap[tag.name] || tagMap[tag.name].confidence < tag.confidence) tagMap[tag.name] = tag;
|
||||
}
|
||||
const deduped: NormalizedTag[] = Object.values(tagMap);
|
||||
|
||||
const { found, missing } = await tagCacheByName.fetch(deduped.map((x) => x.name));
|
||||
let queriedTags: TagWithId[] = [];
|
||||
if (missing.length > 0) {
|
||||
queriedTags = await dbWrite.tag.findMany({
|
||||
where: { name: { in: missing } },
|
||||
select: { id: true, name: true, nsfwLevel: true, type: true },
|
||||
});
|
||||
await tagCacheByName.setMany(queriedTags.map((data) => ({ key: data.name, data })));
|
||||
}
|
||||
const queriedNames = new Set(queriedTags.map((t) => t.name));
|
||||
const tagsToCreate = missing.filter((name) => !queriedNames.has(name));
|
||||
|
||||
let createdTags: TagWithId[] = [];
|
||||
if (tagsToCreate.length > 0) {
|
||||
const tagsToInsert = deduped.filter((x) => tagsToCreate.includes(x.name));
|
||||
|
||||
// Raw SQL bypasses Prisma's @updatedAt stamp, and neither column has a DB
|
||||
// default — both are NOT NULL, so omitting either raises 23502.
|
||||
const now = new Date();
|
||||
const values = tagsToInsert.map(
|
||||
(tag) => Prisma.sql`(${tag.name}, ${now}, ARRAY['Image']::"TagTarget"[])`
|
||||
);
|
||||
|
||||
createdTags = await dbWrite.$queryRaw<TagWithId[]>`
|
||||
INSERT INTO "Tag" (name, "updatedAt", target)
|
||||
VALUES ${Prisma.join(values)}
|
||||
ON CONFLICT (name) DO UPDATE SET name = EXCLUDED.name
|
||||
RETURNING id, name, "nsfwLevel", type
|
||||
`;
|
||||
await tagCacheByName.setMany(createdTags.map((data) => ({ key: data.name, data })));
|
||||
}
|
||||
|
||||
const allTags = [...found.values(), ...queriedTags, ...createdTags]
|
||||
.map((tag) => {
|
||||
const match = normalized.find((x) => x.name === tag.name);
|
||||
if (!match) return null;
|
||||
return { ...tag, source: match.source, confidence: match.confidence };
|
||||
})
|
||||
.filter(isDefined);
|
||||
|
||||
return allTags;
|
||||
}
|
||||
|
||||
// Outcome resolution (audit + moderation rules + persist)
|
||||
// --------------------------------------------------
|
||||
export type ScanOutcome = {
|
||||
ingestion: ImageIngestionStatus;
|
||||
blockedFor: string | null;
|
||||
reviewKey: string | null;
|
||||
/** Present only when the image was audited (i.e. not the `blocked` short-circuit). */
|
||||
audit?: Awaited<ReturnType<typeof auditScanResults>>;
|
||||
};
|
||||
|
||||
/**
|
||||
* Resolve and persist the final state of the image row, returning the bits the
|
||||
* post-update side effects need. Three terminal shapes:
|
||||
* - `blocked` (legacy mediaRating hard block, already written by the caller): only
|
||||
* stamp provenance — re-auditing would recompute nsfwLevel/ingestion and silently
|
||||
* un-block the image. A *prior* block is deliberately not sticky, so a moderator
|
||||
* rescan can clear it. imageScanning never passes `blocked`, so on that path only
|
||||
* `audit.blockedFor` and moderation rules re-block.
|
||||
* - `audit.blockedFor` (prompt TOS/CSAM): block.
|
||||
* - otherwise: Scanned, with moderation rules applied last so a rule Block/Hold
|
||||
* takes precedence over the audit decision.
|
||||
*/
|
||||
export async function resolveScanOutcome({
|
||||
image,
|
||||
blocked,
|
||||
pHash,
|
||||
workflowId,
|
||||
prompt,
|
||||
negativePrompt,
|
||||
log = LEGACY_SCAN_LOG,
|
||||
}: {
|
||||
image: ScanImage;
|
||||
blocked?: { reason: string | null };
|
||||
pHash?: bigint;
|
||||
workflowId: string;
|
||||
prompt?: string;
|
||||
negativePrompt?: string;
|
||||
log?: ScanLog;
|
||||
}): Promise<ScanOutcome> {
|
||||
const updatedAt = new Date();
|
||||
|
||||
if (blocked) {
|
||||
await dbWrite.$executeRaw`
|
||||
UPDATE "Image"
|
||||
SET
|
||||
"updatedAt" = ${updatedAt},
|
||||
"pHash" = COALESCE(${pHash ?? null}, "pHash"),
|
||||
"scanJobs" = jsonb_set(COALESCE("scanJobs", '{}'), '{workflowId}', ${JSON.stringify(
|
||||
workflowId
|
||||
)}::jsonb)
|
||||
WHERE id = ${image.id}
|
||||
`;
|
||||
return {
|
||||
ingestion: ImageIngestionStatus.Blocked,
|
||||
blockedFor: blocked.reason,
|
||||
reviewKey: null,
|
||||
};
|
||||
}
|
||||
|
||||
const audit = await auditScanResults({
|
||||
imageId: image.id,
|
||||
userId: image.userId,
|
||||
prompt,
|
||||
negativePrompt,
|
||||
});
|
||||
let reviewKey = audit.reviewKey ?? null;
|
||||
|
||||
const toUpdate: Prisma.ImageUpdateInput = { updatedAt, pHash };
|
||||
// AI-generation verification is no longer a blocking gate (per operations
|
||||
// 2026-05-11): nsfw images that we couldn't auto-verify as AI used to
|
||||
// land in `Blocked + AiNotVerified`, but the false-positive rate didn't
|
||||
// justify the friction. The remaining `audit.blockedFor` branch still
|
||||
// catches hard violations (TOS / Moderated / CSAM) — everything else
|
||||
// falls through to Scanned.
|
||||
if (audit.blockedFor) {
|
||||
toUpdate.ingestion = ImageIngestionStatus.Blocked;
|
||||
toUpdate.blockedFor = audit.blockedFor;
|
||||
toUpdate.nsfwLevel = NsfwLevel.Blocked;
|
||||
} else {
|
||||
toUpdate.ingestion = ImageIngestionStatus.Scanned;
|
||||
toUpdate.needsReview = reviewKey;
|
||||
toUpdate.minor = audit.minor;
|
||||
toUpdate.poi = audit.poi;
|
||||
toUpdate.blockedFor = null;
|
||||
// Respect a manually-locked nsfw level — never overwrite it from the scan.
|
||||
toUpdate.nsfwLevel = image.nsfwLevelLocked ? image.nsfwLevel : audit.nsfwLevel;
|
||||
|
||||
// scannedAt reassignment: always stamp the first scan; afterwards only
|
||||
// re-stamp recent (<1 week old) non-Rescan images that haven't opted out
|
||||
// via metadata.skipScannedAtReassignment. Older/rescanned images keep
|
||||
// their original scannedAt.
|
||||
const now = new Date();
|
||||
if (!image.scannedAt) toUpdate.scannedAt = now;
|
||||
else if (
|
||||
!(image.metadata as any)?.skipScannedAtReassignment &&
|
||||
image.ingestion !== 'Rescan' &&
|
||||
new Date(image.createdAt).getTime() >= decreaseDate(now, 7, 'days').getTime()
|
||||
)
|
||||
toUpdate.scannedAt = now;
|
||||
else toUpdate.scannedAt = image.scannedAt;
|
||||
}
|
||||
|
||||
// Moderation rules can block the image, hold it for review, or annotate its
|
||||
// metadata with the matched rule. Applied after the scan audit so a Block/Hold
|
||||
// takes precedence over the audit's own decision.
|
||||
const modRule = await evaluateImageModRules(image, audit.tags);
|
||||
let metadataUpdate: Record<string, any> | undefined;
|
||||
if (modRule) {
|
||||
metadataUpdate = modRule.metadata;
|
||||
if (modRule.ingestion) toUpdate.ingestion = modRule.ingestion;
|
||||
if (modRule.nsfwLevel != null) toUpdate.nsfwLevel = modRule.nsfwLevel;
|
||||
if (modRule.blockedFor) toUpdate.blockedFor = modRule.blockedFor;
|
||||
if (modRule.needsReview !== undefined) {
|
||||
toUpdate.needsReview = modRule.needsReview;
|
||||
if (typeof modRule.needsReview === 'string') reviewKey = modRule.needsReview;
|
||||
}
|
||||
// Notify the user only when a rule auto-blocked the image (not on Hold).
|
||||
if (modRule.ingestion === ImageIngestionStatus.Blocked) {
|
||||
await notifyImageAutoBlocked({
|
||||
imageId: image.id,
|
||||
userId: image.userId,
|
||||
reason: modRule.ruleReason,
|
||||
log,
|
||||
});
|
||||
}
|
||||
}
|
||||
|
||||
await dbWrite.$executeRaw`
|
||||
UPDATE "Image"
|
||||
SET
|
||||
"updatedAt" = ${toUpdate.updatedAt},
|
||||
"pHash" = ${pHash ?? null},
|
||||
"ingestion" = ${toUpdate.ingestion as string}::"ImageIngestionStatus",
|
||||
"blockedFor" = ${(toUpdate.blockedFor as string) ?? null},
|
||||
"nsfwLevel" = ${toUpdate.nsfwLevel as number},
|
||||
"needsReview" = ${(toUpdate.needsReview as string) ?? null},
|
||||
"minor" = ${(toUpdate.minor as boolean) ?? false},
|
||||
"poi" = ${(toUpdate.poi as boolean) ?? false},
|
||||
"scannedAt" = ${(toUpdate.scannedAt as Date) ?? null},
|
||||
"metadata" = COALESCE(${
|
||||
metadataUpdate ? JSON.stringify(metadataUpdate) : null
|
||||
}::jsonb, "metadata"),
|
||||
"scanJobs" = jsonb_set(COALESCE("scanJobs", '{}'), '{workflowId}', ${JSON.stringify(
|
||||
workflowId
|
||||
)}::jsonb)
|
||||
WHERE id = ${image.id}
|
||||
`;
|
||||
|
||||
return {
|
||||
ingestion: toUpdate.ingestion as ImageIngestionStatus,
|
||||
blockedFor: (toUpdate.blockedFor as string) ?? null,
|
||||
reviewKey,
|
||||
audit,
|
||||
};
|
||||
}
|
||||
|
||||
async function auditScanResults(args: {
|
||||
imageId: number;
|
||||
userId: number;
|
||||
prompt?: string;
|
||||
negativePrompt?: string;
|
||||
}) {
|
||||
// Moderator-managed benign phrases (proper nouns / technical terms that coincidentally
|
||||
// contain a detection token) are blanked up front so every downstream check — minor,
|
||||
// poi, blockedFor — sees the same cleaned text. A benign phrase is innocent content, so
|
||||
// it shouldn't feed any detector.
|
||||
const [prompt, negativePrompt] = await Promise.all([
|
||||
stripBenignPhrases(normalizeText(args.prompt), BlocklistType.PromptBenignPhrase),
|
||||
stripBenignPhrases(normalizeText(args.negativePrompt), BlocklistType.NegativeBenignPhrase),
|
||||
]);
|
||||
const tags = await dbWrite.$queryRaw<
|
||||
{ id: number; name: string; type: TagType; nsfwLevel: number; confidence: number }[]
|
||||
>`
|
||||
SELECT t.id, t.name, t."nsfwLevel", toi.confidence
|
||||
FROM "TagsOnImageDetails" toi
|
||||
JOIN "Tag" t ON t.id = toi."tagId"
|
||||
WHERE toi."imageId" = ${args.imageId} AND toi.automated AND NOT toi.disabled
|
||||
`;
|
||||
const nsfwLevel = Math.max(...[...tags.map((x) => x.nsfwLevel), 0]);
|
||||
const nsfw = nsfwLevel > sfwBrowsingLevelsFlag;
|
||||
const minorTags = tags.filter((tag) => tagsNeedingReview.includes(tag.name.toLowerCase()));
|
||||
const poiTags = tags.filter((tag) => poiWords.includes(tag.name.toLowerCase()));
|
||||
const reviewTags = [
|
||||
...tags.filter((tag) => tag.nsfwLevel === NsfwLevel.Blocked),
|
||||
...getConditionalTagsForReview(tags, nsfwLevel),
|
||||
];
|
||||
const adultTags = tags.filter((tag) => tag.name === 'adult');
|
||||
const cartoonTags = tags.filter((tag) => styleTags.includes(tag.name));
|
||||
|
||||
const tagReview = reviewTags.length > 0;
|
||||
let poiReview = poiTags.length > 0;
|
||||
let minorReview =
|
||||
minorTags.length > 0 && adultTags.length === 0 && (cartoonTags.length === 0 || nsfw);
|
||||
let newUserReview = false;
|
||||
|
||||
const inappropriate = includesInappropriate({ prompt, negativePrompt }, nsfw);
|
||||
if (inappropriate === 'minor') minorReview = true;
|
||||
if (inappropriate === 'poi') poiReview = true;
|
||||
|
||||
const associatedEntities = await getAssociatedEntities(args.imageId);
|
||||
// Associated poi/minor resources only escalate to review when the image is
|
||||
// nsfw — a sfw image with a poi/minor resource is still flagged (below) but
|
||||
// not queued for moderator review.
|
||||
if (associatedEntities.poi && nsfw) poiReview = true;
|
||||
if (associatedEntities.minor && nsfw) minorReview = true;
|
||||
|
||||
if (!minorReview && !poiReview && !tagReview && nsfw) {
|
||||
newUserReview = await getIsNewUser(args.userId);
|
||||
}
|
||||
|
||||
const minor = minorTags.length > 0 || !!associatedEntities.minor;
|
||||
const poi = poiTags.length > 0 || !!associatedEntities.poi || (!!prompt && !!includesPoi(prompt));
|
||||
|
||||
let reviewKey: string | undefined;
|
||||
if (poiReview) reviewKey = 'poi';
|
||||
else if (minorReview) reviewKey = 'minor';
|
||||
else if (tagReview) reviewKey = 'tag';
|
||||
else if (newUserReview) reviewKey = 'newUser';
|
||||
|
||||
let blockedFor: string | undefined;
|
||||
if (nsfw && prompt) {
|
||||
const auditResult = auditMetaData({ prompt }, nsfw);
|
||||
if (!auditResult.success)
|
||||
blockedFor = auditResult.blockedFor.join(',') ?? 'Failed audit, no explanation';
|
||||
}
|
||||
|
||||
return {
|
||||
tags,
|
||||
nsfwLevel,
|
||||
nsfw,
|
||||
minorTags,
|
||||
poiTags,
|
||||
reviewTags,
|
||||
tagReview,
|
||||
poiReview,
|
||||
minorReview,
|
||||
newUserReview,
|
||||
minor,
|
||||
poi,
|
||||
reviewKey,
|
||||
blockedFor,
|
||||
};
|
||||
}
|
||||
|
||||
async function getAssociatedEntities(imageId: number) {
|
||||
const [result] = await dbWrite.$queryRaw<
|
||||
{ poi: boolean; minor: boolean; hasResource: boolean }[]
|
||||
>`
|
||||
WITH to_check AS (
|
||||
-- Check based on associated resources
|
||||
SELECT
|
||||
SUM(IIF(m.poi, 1, 0)) > 0 "poi",
|
||||
SUM(IIF(m.minor, 1, 0)) > 0 "minor",
|
||||
true "hasResource"
|
||||
FROM "ImageResourceNew" ir
|
||||
JOIN "ModelVersion" mv ON ir."modelVersionId" = mv.id
|
||||
JOIN "Model" m ON m.id = mv."modelId"
|
||||
WHERE ir."imageId" = ${imageId}
|
||||
UNION
|
||||
-- Check based on associated bounties
|
||||
SELECT
|
||||
SUM(IIF(b.poi, 1, 0)) > 0 "poi",
|
||||
false "minor",
|
||||
false "hasResource"
|
||||
FROM "Image" i
|
||||
JOIN "ImageConnection" ic ON ic."imageId" = i.id
|
||||
JOIN "Bounty" b ON ic."entityType" = 'Bounty' AND b.id = ic."entityId"
|
||||
WHERE ic."imageId" = ${imageId}
|
||||
UNION
|
||||
-- Check based on associated bounty entries
|
||||
SELECT
|
||||
SUM(IIF(b.poi, 1, 0)) > 0 "poi",
|
||||
false "minor",
|
||||
false "hasResource"
|
||||
FROM "Image" i
|
||||
JOIN "ImageConnection" ic ON ic."imageId" = i.id
|
||||
JOIN "BountyEntry" be ON ic."entityType" = 'BountyEntry' AND be.id = ic."entityId"
|
||||
JOIN "Bounty" b ON b.id = be."bountyId"
|
||||
WHERE ic."imageId" = ${imageId}
|
||||
)
|
||||
SELECT bool_or(poi) "poi", bool_or(minor) "minor", bool_or("hasResource") "hasResource" FROM to_check;
|
||||
`;
|
||||
|
||||
return result;
|
||||
}
|
||||
|
||||
async function getIsNewUser(userId: number) {
|
||||
const [{ isNewUser }] =
|
||||
(await dbWrite.$queryRaw<{ isNewUser: boolean }[]>`
|
||||
SELECT is_new_user(CAST(${userId} AS INT)) "isNewUser";
|
||||
`) ?? [];
|
||||
return isNewUser;
|
||||
}
|
||||
|
||||
// Moderation rules
|
||||
// --------------------------------------------------
|
||||
/**
|
||||
* Evaluate the image moderation rules and return the row changes they imply
|
||||
* (block / hold / metadata annotation). Pure of side effects — the caller is
|
||||
* responsible for persisting the changes and, on a Block, calling
|
||||
* `notifyImageAutoBlocked`.
|
||||
*/
|
||||
async function evaluateImageModRules(
|
||||
image: { meta: Prisma.JsonValue; metadata: Prisma.JsonValue },
|
||||
tags: { name: string }[]
|
||||
) {
|
||||
const imageModRules = await getImagesModRules();
|
||||
if (!imageModRules.length) return;
|
||||
|
||||
const tagNames = tags.map((x) => x.name);
|
||||
const meta = (image.meta ?? {}) as Prisma.JsonObject;
|
||||
const appliedRule = evaluateRules(imageModRules, { ...meta, tags: tagNames });
|
||||
if (!appliedRule || appliedRule.action === ModerationRuleAction.Approve) return;
|
||||
|
||||
const result: {
|
||||
metadata: Record<string, any>;
|
||||
ruleReason?: string | null;
|
||||
ingestion?: ImageIngestionStatus;
|
||||
nsfwLevel?: NsfwLevel;
|
||||
blockedFor?: string;
|
||||
needsReview?: string | null;
|
||||
} = {
|
||||
metadata: {
|
||||
...((image.metadata ?? {}) as Record<string, any>),
|
||||
ruleId: appliedRule.id,
|
||||
ruleReason: appliedRule.reason,
|
||||
},
|
||||
ruleReason: appliedRule.reason,
|
||||
};
|
||||
|
||||
if (appliedRule.action === ModerationRuleAction.Block) {
|
||||
result.ingestion = ImageIngestionStatus.Blocked;
|
||||
result.nsfwLevel = NsfwLevel.Blocked;
|
||||
result.blockedFor = BlockedReason.Moderated;
|
||||
result.needsReview = null;
|
||||
} else if (appliedRule.action === ModerationRuleAction.Hold) {
|
||||
result.needsReview = 'modRule';
|
||||
}
|
||||
|
||||
return result;
|
||||
}
|
||||
|
||||
async function notifyImageAutoBlocked({
|
||||
imageId,
|
||||
userId,
|
||||
reason,
|
||||
log,
|
||||
}: {
|
||||
imageId: number;
|
||||
userId: number;
|
||||
reason?: string | null;
|
||||
log: ScanLog;
|
||||
}) {
|
||||
await createNotification({
|
||||
category: NotificationCategory.System,
|
||||
key: `image-block:${imageId}`,
|
||||
type: 'system-message',
|
||||
userId,
|
||||
details: {
|
||||
message: `One of your images has been blocked due to a moderation rule violation${
|
||||
reason ? ` by the following reason: ${reason}` : ''
|
||||
}. If you believe this is a mistake, you can appeal this decision.`,
|
||||
url: `/images/${imageId}`,
|
||||
},
|
||||
}).catch((error) =>
|
||||
logToAxiom({
|
||||
name: log.name,
|
||||
type: 'error',
|
||||
message: 'Could not create notification when blocking image',
|
||||
data: {
|
||||
imageId,
|
||||
error: error.message,
|
||||
cause: error.cause,
|
||||
stack: error.stack,
|
||||
},
|
||||
})
|
||||
);
|
||||
}
|
||||
|
||||
// Post-update side effects
|
||||
// --------------------------------------------------
|
||||
export async function applyIngestionSideEffects({
|
||||
image,
|
||||
outcome,
|
||||
}: {
|
||||
image: ScanImage;
|
||||
outcome: ScanOutcome;
|
||||
}) {
|
||||
// handle blocked image updates
|
||||
if (outcome.ingestion === ImageIngestionStatus.Blocked) {
|
||||
await queueImageSearchIndexUpdate({
|
||||
ids: [image.id],
|
||||
action: SearchIndexUpdateQueueAction.Delete,
|
||||
});
|
||||
// A previously-cached Blocked image can still satisfy the showcase query
|
||||
// filters (needsReview IS NULL, nsfwLevel != 0) so drop it from the showcase.
|
||||
if (image.postId) await bustCachesForPosts(image.postId);
|
||||
await updateModel3DNsfwLevelForThumbnailImage({ imageId: image.id, postId: image.postId });
|
||||
// If this image belongs to a comic panel, the parent project may
|
||||
// have been search-indexed under the old (unblocked) state. Re-queue
|
||||
// it so the next index pass re-evaluates visibility against the
|
||||
// moderation gates in `comics.search-index.ts:WHERE`.
|
||||
await queueComicsForPanelImage(image.id);
|
||||
return;
|
||||
}
|
||||
|
||||
// handle scanned image updates
|
||||
if (outcome.ingestion === ImageIngestionStatus.Scanned) {
|
||||
// Scanning is what makes an already-published image countable. Bust rather
|
||||
// than refresh: this fires once per image, so a re-query here would be N
|
||||
// identical counts for an N-image post.
|
||||
await userImageVideoCountCaches.bust(image.userId);
|
||||
await tagIdsForImagesCache.refresh(image.id);
|
||||
if (
|
||||
typeof image.metadata === 'object' &&
|
||||
(image.metadata as MediaMetadata | undefined)?.profilePicture
|
||||
) {
|
||||
await deleteUserProfilePictureCache(image.userId);
|
||||
}
|
||||
|
||||
if (image.postId) {
|
||||
await updatePostNsfwLevel(image.postId);
|
||||
// Without this, the showcase cache stays empty until its 24h TTL for any model version whose images hadn't scanned yet on first read.
|
||||
await bustCachesForPosts(image.postId);
|
||||
}
|
||||
await updateModel3DNsfwLevelForThumbnailImage({ imageId: image.id, postId: image.postId });
|
||||
await updateComicNsfwLevelsForImage(image.id);
|
||||
// Refresh the comic project in the search index — even on a clean
|
||||
// Scanned, `needsReview` may have been set, which the index treats
|
||||
// as a visibility gate.
|
||||
await queueComicsForPanelImage(image.id);
|
||||
|
||||
await queueImageSearchIndexUpdate({
|
||||
ids: [image.id],
|
||||
action: SearchIndexUpdateQueueAction.Update,
|
||||
});
|
||||
|
||||
const { audit, reviewKey } = outcome;
|
||||
if (audit) {
|
||||
const tagsForReview = [...audit.poiTags, ...audit.minorTags, ...audit.reviewTags];
|
||||
// Only persist review-tags when the image is actually queued for review
|
||||
// (matches the legacy `if (reviewKey)` gate).
|
||||
if (reviewKey && tagsForReview.length > 0) {
|
||||
await createImageTagsForReview({
|
||||
imageId: image.id,
|
||||
tagIds: tagsForReview.map((x) => x.id),
|
||||
});
|
||||
}
|
||||
|
||||
if (!reviewKey && image.type === 'image') {
|
||||
await addToNewOrderQueue({ imageId: image.id, nsfw: audit.nsfw });
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
const KONO_NSFW_SAMPLING_RATE = 0.3; // 30%
|
||||
async function addToNewOrderQueue({ imageId, nsfw }: { imageId: number; nsfw: boolean }) {
|
||||
let shouldAddToQueue = true;
|
||||
let priority: 1 | 2 | 3 = 1;
|
||||
const rankType = NewOrderRankType.Knight;
|
||||
if (nsfw) {
|
||||
priority = 2;
|
||||
shouldAddToQueue = Math.random() < KONO_NSFW_SAMPLING_RATE;
|
||||
}
|
||||
if (shouldAddToQueue) {
|
||||
await addImageToQueue({
|
||||
imageIds: [imageId],
|
||||
rankType,
|
||||
priority,
|
||||
});
|
||||
}
|
||||
}
|
||||
File diff suppressed because it is too large
Load Diff
@@ -1,25 +0,0 @@
|
||||
/**
|
||||
* Pure decision helper for the `image-scanner-new` sysRedis kill-switch.
|
||||
*
|
||||
* Lives in its own tiny module so the Buffer-coercion + seed-only-on-unset
|
||||
* behavior is unit-testable without importing the ~7.9K-line image.service
|
||||
* (which drags in Prisma/env/auth at load time and can't load under vitest).
|
||||
*
|
||||
* sysRedis.get is typed `string` but the HA/Sentinel client returns a Buffer
|
||||
* for BLOB_STRING replies, matching none of the four literals → pre-fix
|
||||
* `isImageScannerNewEnabled` fell through and DESTRUCTIVELY overwrote the
|
||||
* operator's '1' with 'false' (then returned false). Coerce once before
|
||||
* comparing. See PR #2697/#2700 for the canonical Buffer-vs-string regression.
|
||||
*
|
||||
* Returns:
|
||||
* - `true` for '1' / 'true'
|
||||
* - `false` for '0' / 'false'
|
||||
* - `null` for a genuinely-unset/unknown key — the ONLY case in which the
|
||||
* caller should default-seed 'false'.
|
||||
*/
|
||||
export function parseScannerFlag(raw: unknown): boolean | null {
|
||||
const value = Buffer.isBuffer(raw) ? raw.toString('utf8') : raw;
|
||||
if (value === '1' || value === 'true') return true;
|
||||
if (value === '0' || value === 'false') return false;
|
||||
return null;
|
||||
}
|
||||
@@ -0,0 +1,310 @@
|
||||
import type {
|
||||
ImageScanningJointAgeDetection,
|
||||
ImageScanningOutput,
|
||||
NsfwLevel,
|
||||
} from '@civitai/orchestration-client';
|
||||
import { orchestratorNsfwLevelMap } from '~/shared/constants/browsingLevel.constants';
|
||||
import { logToAxiom } from '~/server/logging/client';
|
||||
import {
|
||||
applyIngestionSideEffects,
|
||||
buildAndInsertScanTags,
|
||||
extractFailedSteps,
|
||||
loadImageForScan,
|
||||
logPerceptualHashMatch,
|
||||
markImageScanError,
|
||||
readJobFailureReason,
|
||||
resolveScanOutcome,
|
||||
sendIngestionSignal,
|
||||
type ScanLog,
|
||||
} from '~/server/services/image-scan-pipeline';
|
||||
import { removeImageScanJobQueue } from '~/server/services/job-queue.service';
|
||||
import { computePerceptualHash } from '~/server/services/orchestrator/orchestrator.service';
|
||||
import { recordImageScanningResult } from '~/server/services/scanner-audit.service';
|
||||
import { fanOutArticleImageUpdates } from '~/server/utils/webhook-debounce';
|
||||
import { ImageIngestionStatus } from '~/shared/utils/prisma/enums';
|
||||
|
||||
type RawRepeatStep = {
|
||||
$type?: string;
|
||||
input?: { template?: { $type?: string } };
|
||||
output?: { steps?: Array<{ $type?: string }> };
|
||||
};
|
||||
|
||||
/** Whether a workflow was submitted with imageScanning, whatever state its steps ended in. */
|
||||
export function isImageScanningWorkflow(steps: unknown[]) {
|
||||
return (steps as RawRepeatStep[]).some(
|
||||
(step) =>
|
||||
step.$type === 'imageScanning' ||
|
||||
(step.$type === 'repeat' &&
|
||||
(step.input?.template?.$type === 'imageScanning' ||
|
||||
step.output?.steps?.[0]?.$type === 'imageScanning'))
|
||||
);
|
||||
}
|
||||
|
||||
type Recognition = { label: string; score: number };
|
||||
|
||||
/** One image's scan, or a video's frames combined into one. */
|
||||
export type ImageScanningResult = {
|
||||
nsfwLevel: NsfwLevel;
|
||||
tags: Record<string, number>;
|
||||
csam: boolean | null;
|
||||
ageDetections: ImageScanningJointAgeDetection[];
|
||||
minorDetected: boolean;
|
||||
aiRecognition?: Recognition;
|
||||
animeRecognition?: Recognition;
|
||||
perceptualHash?: string;
|
||||
};
|
||||
|
||||
type RawStep = { $type?: string; name?: string; status?: string; output?: unknown };
|
||||
|
||||
// wdTagging only ever returned general tags; the other categories would write tags no image had before.
|
||||
const INGESTED_TAG_CATEGORIES = new Set(['general']);
|
||||
|
||||
function readScan(output: ImageScanningOutput, workflowId: string): ImageScanningResult {
|
||||
const { tagging, jointAgeClassification } = output;
|
||||
// An unmapped level ('na', or one added upstream) would write a rating with no NSFW level.
|
||||
// Not "invalid media…": classifyImageScanFailure reads that as permanent, and this is recoverable.
|
||||
if (orchestratorNsfwLevelMap[output.nsfwLevel] === undefined)
|
||||
throw new Error(`media rating unavailable (${output.nsfwLevel}) for workflow: ${workflowId}`);
|
||||
if (!tagging?.ran)
|
||||
throw new Error(`Incomplete workflow: ${workflowId}. Tagging did not run (${tagging?.status})`);
|
||||
|
||||
const tags: Record<string, number> = {};
|
||||
for (const { tag, category, score } of tagging.tags) {
|
||||
if (INGESTED_TAG_CATEGORIES.has(category)) tags[tag] = Math.max(score, tags[tag] ?? 0);
|
||||
}
|
||||
|
||||
return {
|
||||
nsfwLevel: output.nsfwLevel,
|
||||
tags,
|
||||
csam: output.csam ?? null,
|
||||
ageDetections: jointAgeClassification?.detections ?? [],
|
||||
minorDetected: !!jointAgeClassification?.minorDetected,
|
||||
aiRecognition: output.aiRecognition,
|
||||
animeRecognition: output.animeRecognition,
|
||||
};
|
||||
}
|
||||
|
||||
const higher = (a?: Recognition, b?: Recognition) => (!a || (b && b.score > a.score) ? b : a);
|
||||
|
||||
// Every result is kept (a max or an OR per field), so a frame can only raise what the video reports.
|
||||
function combineFrames(frames: ImageScanningResult[]): ImageScanningResult {
|
||||
return frames.reduce((acc, frame) => {
|
||||
const tags = { ...acc.tags };
|
||||
for (const [tag, score] of Object.entries(frame.tags))
|
||||
tags[tag] = Math.max(score, tags[tag] ?? 0);
|
||||
return {
|
||||
nsfwLevel:
|
||||
orchestratorNsfwLevelMap[frame.nsfwLevel] > orchestratorNsfwLevelMap[acc.nsfwLevel]
|
||||
? frame.nsfwLevel
|
||||
: acc.nsfwLevel,
|
||||
tags,
|
||||
csam: acc.csam || frame.csam ? true : acc.csam ?? frame.csam,
|
||||
ageDetections: [...acc.ageDetections, ...frame.ageDetections],
|
||||
minorDetected: acc.minorDetected || frame.minorDetected,
|
||||
aiRecognition: higher(acc.aiRecognition, frame.aiRecognition),
|
||||
animeRecognition: higher(acc.animeRecognition, frame.animeRecognition),
|
||||
};
|
||||
});
|
||||
}
|
||||
|
||||
/**
|
||||
* Reads a succeeded imageScanning ingestion workflow: one `imageScanning` step for an image,
|
||||
* or a `repeat` of them over extracted frames for a video, plus the `mediaHash` step.
|
||||
* Throws when the workflow carries nothing usable.
|
||||
*/
|
||||
export function parseImageScanningSteps(workflowSteps: unknown[], workflowId: string) {
|
||||
const steps = workflowSteps as RawStep[];
|
||||
const perceptualHash = (
|
||||
steps.find((step) => step.$type === 'mediaHash')?.output as
|
||||
| { hashes?: { perceptual?: string } }
|
||||
| undefined
|
||||
)?.hashes?.perceptual;
|
||||
|
||||
const single = steps.find((step) => step.$type === 'imageScanning')?.output;
|
||||
if (single) return { ...readScan(single as ImageScanningOutput, workflowId), perceptualHash };
|
||||
|
||||
const frames = (
|
||||
steps.find(
|
||||
(step) =>
|
||||
step.$type === 'repeat' &&
|
||||
(step.output as { steps?: RawStep[] } | undefined)?.steps?.[0]?.$type === 'imageScanning'
|
||||
)?.output as { steps: RawStep[] } | undefined
|
||||
)?.steps;
|
||||
if (!frames?.length || !frames.every((frame) => frame.output))
|
||||
throw new Error(`Incomplete workflow: ${workflowId}. Missing imageScanning output`);
|
||||
|
||||
return {
|
||||
...combineFrames(
|
||||
frames.map((frame) => readScan(frame.output as ImageScanningOutput, workflowId))
|
||||
),
|
||||
perceptualHash,
|
||||
};
|
||||
}
|
||||
|
||||
export const IMAGE_SCANNING_LOG: ScanLog = {
|
||||
name: 'image-scanning-result',
|
||||
source: 'image-scanning-result.service',
|
||||
};
|
||||
|
||||
type ScanErrorInput = Parameters<typeof markImageScanError>[0];
|
||||
|
||||
async function failScan(
|
||||
input: ScanErrorInput & {
|
||||
articleImageScanning: boolean;
|
||||
logType: 'warning' | 'error';
|
||||
message: string;
|
||||
stack?: string;
|
||||
}
|
||||
) {
|
||||
const { articleImageScanning, logType, message, stack, ...error } = input;
|
||||
const { retryCount, mediaType, userId, failureClass } = await markImageScanError(error);
|
||||
logToAxiom(
|
||||
{
|
||||
name: IMAGE_SCANNING_LOG.name,
|
||||
type: logType,
|
||||
message,
|
||||
source: IMAGE_SCANNING_LOG.source,
|
||||
stack,
|
||||
...error,
|
||||
failureClass,
|
||||
mediaType,
|
||||
retryCount,
|
||||
},
|
||||
'webhooks'
|
||||
).catch(() => null);
|
||||
await sendIngestionSignal({
|
||||
imageId: error.imageId,
|
||||
userId,
|
||||
ingestion: ImageIngestionStatus.Error,
|
||||
log: IMAGE_SCANNING_LOG,
|
||||
});
|
||||
if (articleImageScanning) await fanOutArticleImageUpdates(error.imageId);
|
||||
}
|
||||
|
||||
/** Core image scan processing for imageScanning workflows. */
|
||||
export async function processImageScanningWorkflow({
|
||||
workflowId,
|
||||
status,
|
||||
steps,
|
||||
imageId,
|
||||
articleImageScanning = false,
|
||||
startedAt,
|
||||
completedAt,
|
||||
}: {
|
||||
workflowId: string;
|
||||
status: string;
|
||||
steps: unknown[];
|
||||
imageId: number;
|
||||
articleImageScanning?: boolean;
|
||||
startedAt?: Date | string | null;
|
||||
completedAt?: Date | string | null;
|
||||
}) {
|
||||
if (status !== 'succeeded') {
|
||||
await failScan({
|
||||
workflowId,
|
||||
imageId,
|
||||
status,
|
||||
failureType:
|
||||
status === 'expired' ? 'expired' : status === 'canceled' ? 'canceled' : 'workflow-failed',
|
||||
failedSteps: extractFailedSteps(steps),
|
||||
reason: await readJobFailureReason(workflowId),
|
||||
articleImageScanning,
|
||||
logType: 'warning',
|
||||
message: `workflow not succeeded: ${status}`,
|
||||
});
|
||||
return;
|
||||
}
|
||||
|
||||
let scan: ReturnType<typeof parseImageScanningSteps>;
|
||||
try {
|
||||
scan = parseImageScanningSteps(steps, workflowId);
|
||||
} catch (error) {
|
||||
await failScan({
|
||||
workflowId,
|
||||
imageId,
|
||||
status,
|
||||
failureType: 'unusable-result',
|
||||
failedSteps: extractFailedSteps(steps),
|
||||
reason: error instanceof Error ? error.message : 'Unknown error',
|
||||
articleImageScanning,
|
||||
logType: 'warning',
|
||||
message: 'succeeded workflow had no usable result',
|
||||
});
|
||||
return;
|
||||
}
|
||||
|
||||
let image: Awaited<ReturnType<typeof loadImageForScan>>;
|
||||
let outcome: Awaited<ReturnType<typeof resolveScanOutcome>>;
|
||||
try {
|
||||
const pHash = computePerceptualHash(scan.perceptualHash);
|
||||
if (pHash) await logPerceptualHashMatch({ imageId, pHash, log: IMAGE_SCANNING_LOG });
|
||||
|
||||
image = await loadImageForScan(imageId);
|
||||
const { prompt, negativePrompt } = (image.meta ?? {}) as {
|
||||
prompt?: string;
|
||||
negativePrompt?: string;
|
||||
};
|
||||
await buildAndInsertScanTags({
|
||||
imageId,
|
||||
wdTags: scan.tags,
|
||||
ratingLevel: scan.nsfwLevel,
|
||||
prompt,
|
||||
});
|
||||
outcome = await resolveScanOutcome({
|
||||
image,
|
||||
pHash,
|
||||
workflowId,
|
||||
prompt,
|
||||
negativePrompt,
|
||||
log: IMAGE_SCANNING_LOG,
|
||||
});
|
||||
} catch (error) {
|
||||
// Deleted between submit and callback — no row to mark. The route ACKs it.
|
||||
if (error instanceof Error && error.message.startsWith('image not found')) throw error;
|
||||
await failScan({
|
||||
workflowId,
|
||||
imageId,
|
||||
status,
|
||||
failureType: 'processing-failed',
|
||||
failedSteps: extractFailedSteps(steps),
|
||||
reason: error instanceof Error ? error.message : 'Unknown error',
|
||||
articleImageScanning,
|
||||
logType: 'error',
|
||||
message: 'failed to process a succeeded workflow',
|
||||
stack: error instanceof Error ? error.stack : undefined,
|
||||
});
|
||||
return;
|
||||
}
|
||||
|
||||
// The verdict is persisted; throwing now would 400 a finished webhook and make the
|
||||
// orchestrator re-run it.
|
||||
try {
|
||||
await removeImageScanJobQueue([imageId]);
|
||||
await recordImageScanningResult({ workflowId, imageId, scan, startedAt, completedAt });
|
||||
if (articleImageScanning) await fanOutArticleImageUpdates(imageId);
|
||||
await applyIngestionSideEffects({ image, outcome });
|
||||
} catch (error) {
|
||||
logToAxiom(
|
||||
{
|
||||
name: IMAGE_SCANNING_LOG.name,
|
||||
type: 'error',
|
||||
message: 'side effects failed after the scan verdict was persisted',
|
||||
source: IMAGE_SCANNING_LOG.source,
|
||||
stack: error instanceof Error ? error.stack : undefined,
|
||||
reason: error instanceof Error ? error.message : 'Unknown error',
|
||||
imageId,
|
||||
workflowId,
|
||||
ingestion: outcome.ingestion,
|
||||
},
|
||||
'webhooks'
|
||||
).catch(() => null);
|
||||
}
|
||||
|
||||
await sendIngestionSignal({
|
||||
imageId,
|
||||
userId: image.userId,
|
||||
ingestion: outcome.ingestion,
|
||||
blockedFor: outcome.blockedFor,
|
||||
log: IMAGE_SCANNING_LOG,
|
||||
});
|
||||
}
|
||||
@@ -37,7 +37,6 @@ import {
|
||||
import { imageReviewedSql } from '~/server/common/image-visibility';
|
||||
import {
|
||||
BlockedReason,
|
||||
ImageScanType,
|
||||
ImageSort,
|
||||
NotificationCategory,
|
||||
NsfwLevel,
|
||||
@@ -140,11 +139,7 @@ import type {
|
||||
} from '~/server/schema/image.schema';
|
||||
import { imageMetaOutput, ingestImageSchema } from '~/server/schema/image.schema';
|
||||
import type { ImageMetadata, VideoMetadata } from '~/server/schema/media.schema';
|
||||
import {
|
||||
articlesSearchIndex,
|
||||
imagesMetricsSearchIndex,
|
||||
imagesSearchIndex,
|
||||
} from '~/server/search-index';
|
||||
import { imagesMetricsSearchIndex, imagesSearchIndex } from '~/server/search-index';
|
||||
import type {
|
||||
ImageMetricsSearchIndexRecord,
|
||||
MetricsImageFilterableAttribute,
|
||||
@@ -172,7 +167,6 @@ import {
|
||||
} from '~/server/services/model3d.service';
|
||||
import { addImageToQueue } from '~/server/services/games/new-order.service';
|
||||
import { upsertImageFlag } from '~/server/services/image-flag.service';
|
||||
import { parseScannerFlag } from '~/server/services/image-scanner-flag';
|
||||
import {
|
||||
deleteImagTagsForReviewByImageIds,
|
||||
getImagTagsForReviewByImageIds,
|
||||
@@ -203,7 +197,6 @@ import {
|
||||
throwInternalServerError,
|
||||
throwNotFoundError,
|
||||
} from '~/server/utils/errorHandling';
|
||||
import { fetchTimeoutSignal } from '~/server/utils/fetch-timeout';
|
||||
import type { RuleDefinition } from '~/server/utils/mod-rules';
|
||||
import { getCursor } from '~/server/utils/pagination-helpers';
|
||||
import {
|
||||
@@ -267,7 +260,10 @@ import type {
|
||||
} from '../../../event-engine-common/types/package-stubs';
|
||||
import type { FeedQueryInput } from '../../../event-engine-common/feeds/types';
|
||||
import type { ImageQueryInput } from '../../../event-engine-common/types/image-feed-types';
|
||||
import { createImageIngestionRequest } from '~/server/services/orchestrator/orchestrator.service';
|
||||
import {
|
||||
createImageIngestionRequest,
|
||||
imageIngestionLogName,
|
||||
} from '~/server/services/orchestrator/orchestrator.service';
|
||||
import { getGenerationDisplayKeys } from '~/server/services/orchestrator/legacy-metadata-mapper';
|
||||
import {
|
||||
sanitizeProvenance,
|
||||
@@ -1154,29 +1150,6 @@ export const getImageById = async ({ id }: GetByIdInput) => {
|
||||
});
|
||||
};
|
||||
|
||||
/**
|
||||
* Runtime toggle for the new image ingestion path (createImageIngestionRequest
|
||||
* with the expanded mediaRating step). Reads from Redis so ops can flip
|
||||
* without a deploy. Accepts '1'/'true' / '0'/'false' as string values.
|
||||
*
|
||||
* If the key doesn't exist (first request after deploy), seeds it to 'false'
|
||||
* so the toggle is discoverable in Redis and explicitly off by default.
|
||||
* Operators set the key to '1' to enable.
|
||||
*/
|
||||
async function isImageScannerNewEnabled(): Promise<boolean> {
|
||||
// The HA/Sentinel sysRedis returns a Buffer for BLOB_STRING replies, which
|
||||
// matched none of the literals pre-fix → fell through and destructively
|
||||
// overwrote the operator's '1' with 'false'. parseScannerFlag coerces the
|
||||
// Buffer first and returns null ONLY for a genuinely-unset/unknown key, so
|
||||
// the seed below now fires only in its intended default-seeding case.
|
||||
// See PR #2697/#2700 for the canonical Buffer-vs-string regression.
|
||||
const raw = await sysRedis.get(REDIS_SYS_KEYS.SYSTEM.IMAGE_SCANNER_NEW);
|
||||
const parsed = parseScannerFlag(raw);
|
||||
if (parsed !== null) return parsed;
|
||||
await sysRedis.set(REDIS_SYS_KEYS.SYSTEM.IMAGE_SCANNER_NEW, 'false');
|
||||
return false;
|
||||
}
|
||||
|
||||
export const ingestImageById = async ({ id }: GetByIdInput) => {
|
||||
const images = await dbWrite.$queryRaw<IngestImageInput[]>`
|
||||
SELECT id, url, type, width, height, meta->>'prompt' as prompt
|
||||
@@ -1198,16 +1171,6 @@ export const ingestImageById = async ({ id }: GetByIdInput) => {
|
||||
return await ingestImage({ image: images[0] });
|
||||
};
|
||||
|
||||
// const scanner = env.EXTERNAL_IMAGE_SCANNER;
|
||||
// const clavataScan = env.CLAVATA_SCAN;
|
||||
export const imageScanTypes: ImageScanType[] = [
|
||||
ImageScanType.WD14,
|
||||
// ImageScanType.Hash,
|
||||
// ImageScanType.Clavata,
|
||||
// ImageScanType.Hive,
|
||||
ImageScanType.SpineRating,
|
||||
];
|
||||
|
||||
function extractSubmitErrorMessage(error: unknown): string | null {
|
||||
if (!error) return null;
|
||||
if (typeof error === 'string') return error;
|
||||
@@ -1292,233 +1255,64 @@ export const ingestImage = async ({
|
||||
const scanRequestedAt = new Date();
|
||||
const dbClient = tx ?? dbWrite;
|
||||
|
||||
// if (!isProd || !env.IMAGE_SCANNING_ENDPOINT) {
|
||||
// console.log('skipping image ingestion');
|
||||
// const updated = await dbClient.image.update({
|
||||
// where: { id: image.id },
|
||||
// select: { postId: true },
|
||||
// data: {
|
||||
// scanRequestedAt,
|
||||
// scannedAt: scanRequestedAt,
|
||||
// ingestion: ImageIngestionStatus.Scanned,
|
||||
// nsfwLevel: NsfwLevel.PG,
|
||||
// },
|
||||
// });
|
||||
|
||||
// // Update post NSFW level
|
||||
// if (updated.postId) await updatePostNsfwLevel(updated.postId);
|
||||
|
||||
// return true;
|
||||
// }
|
||||
|
||||
const parsedImage = ingestImageSchema.safeParse(image);
|
||||
if (!parsedImage.success) throw new Error('Failed to parse image data');
|
||||
|
||||
const { url, id, type, width, height } = parsedImage.data;
|
||||
const { url, id, type } = parsedImage.data;
|
||||
|
||||
const callbackUrl =
|
||||
env.IMAGE_SCANNING_CALLBACK ??
|
||||
`${env.NEXTAUTH_URL}/api/webhooks/image-scan-result?token=${env.WEBHOOK_TOKEN}`;
|
||||
|
||||
if (!image.prompt) {
|
||||
const { prompt } = await dbClient.$queryRaw<{ prompt?: string }>`
|
||||
SELECT meta->>'prompt' as prompt FROM "Image" WHERE id = ${id}
|
||||
`;
|
||||
image.prompt = prompt;
|
||||
}
|
||||
|
||||
if (await isImageScannerNewEnabled()) {
|
||||
const {
|
||||
data: workflowResponse,
|
||||
error: submitError,
|
||||
status: submitStatus,
|
||||
} = await createImageIngestionRequest({
|
||||
imageId: id,
|
||||
url,
|
||||
type,
|
||||
callbackUrl,
|
||||
priority: lowPriority ? 'low' : undefined,
|
||||
});
|
||||
if (!workflowResponse) {
|
||||
imageScanSubmittedCounter.inc({ lane: 'new', result: 'failed' });
|
||||
const failureClass = await markImageScanSubmitFailure({
|
||||
dbClient,
|
||||
imageId: id,
|
||||
status: submitStatus,
|
||||
error: submitError,
|
||||
});
|
||||
// The orchestrator submit already logs the transient failure in
|
||||
// createImageIngestionRequest, but from here it's otherwise a silent
|
||||
// `return false` — surface it at the dispatch layer so the failure is
|
||||
// attributable to a specific image + media type.
|
||||
logToAxiom({
|
||||
name: 'image-ingestion',
|
||||
type: 'error',
|
||||
reason: 'no-workflow-response',
|
||||
failureType: 'send-fail',
|
||||
failureClass,
|
||||
responseStatus: submitStatus,
|
||||
imageId: id,
|
||||
mediaType: type,
|
||||
}).catch(() => null);
|
||||
return false;
|
||||
}
|
||||
const scanJobsJson = JSON.stringify({ workflowId: workflowResponse.id });
|
||||
await dbClient.$executeRaw`
|
||||
UPDATE "Image"
|
||||
SET
|
||||
"scanRequestedAt" = ${scanRequestedAt},
|
||||
"scanJobs" = CASE
|
||||
WHEN "scanJobs" IS NOT NULL AND "scanJobs" ? 'retryCount' THEN
|
||||
${scanJobsJson}::jsonb || jsonb_build_object('retryCount', ("scanJobs"->'retryCount'))
|
||||
ELSE
|
||||
${scanJobsJson}::jsonb
|
||||
END
|
||||
WHERE id = ${id}
|
||||
`;
|
||||
imageScanSubmittedCounter.inc({ lane: 'new', result: 'success' });
|
||||
return true;
|
||||
}
|
||||
|
||||
let scanUrl = `${env.IMAGE_SCANNING_ENDPOINT}/enqueue`;
|
||||
if (lowPriority) scanUrl += '?lowpri=true';
|
||||
|
||||
const response = await fetch(scanUrl, {
|
||||
method: 'POST',
|
||||
headers: { 'Content-Type': 'application/json' },
|
||||
signal: fetchTimeoutSignal(60_000),
|
||||
body: JSON.stringify({
|
||||
imageId: id,
|
||||
imageKey: url,
|
||||
type,
|
||||
width,
|
||||
height,
|
||||
prompt: image.prompt,
|
||||
// wait: true,
|
||||
scans: imageScanTypes,
|
||||
callbackUrl,
|
||||
movieRatingModel: env.IMAGE_SCANNING_MODEL,
|
||||
}),
|
||||
const {
|
||||
data: workflowResponse,
|
||||
error: submitError,
|
||||
status: submitStatus,
|
||||
useImageScanning,
|
||||
} = await createImageIngestionRequest({
|
||||
imageId: id,
|
||||
url,
|
||||
type,
|
||||
callbackUrl,
|
||||
priority: lowPriority ? 'low' : undefined,
|
||||
});
|
||||
if (response.status === 202) {
|
||||
const scanJobs = (await response.json().catch(() => Prisma.JsonNull)) as
|
||||
| { jobId: string }
|
||||
| typeof Prisma.JsonNull;
|
||||
|
||||
// Convert scanJobs to JSON string for raw SQL, preserving existing retryCount if it exists
|
||||
const scanJobsJson = scanJobs === Prisma.JsonNull ? null : JSON.stringify(scanJobs);
|
||||
|
||||
if (scanJobsJson) {
|
||||
await dbClient.$executeRaw`
|
||||
UPDATE "Image"
|
||||
SET
|
||||
"scanRequestedAt" = ${scanRequestedAt},
|
||||
"scanJobs" = CASE
|
||||
WHEN "scanJobs" IS NOT NULL AND "scanJobs" ? 'retryCount' THEN
|
||||
${scanJobsJson}::jsonb || jsonb_build_object('retryCount', ("scanJobs"->'retryCount'))
|
||||
ELSE
|
||||
${scanJobsJson}::jsonb
|
||||
END
|
||||
WHERE id = ${id}
|
||||
`;
|
||||
} else {
|
||||
await dbClient.$executeRaw`
|
||||
UPDATE "Image"
|
||||
SET "scanRequestedAt" = ${scanRequestedAt}
|
||||
WHERE id = ${id}
|
||||
`;
|
||||
}
|
||||
|
||||
imageScanSubmittedCounter.inc({ lane: 'legacy', result: 'success' });
|
||||
return true;
|
||||
} else {
|
||||
await logToAxiom({
|
||||
name: 'image-ingestion',
|
||||
type: 'error',
|
||||
const lane = useImageScanning ? 'imageScanning' : 'new';
|
||||
if (!workflowResponse) {
|
||||
imageScanSubmittedCounter.inc({ lane, result: 'failed' });
|
||||
const failureClass = await markImageScanSubmitFailure({
|
||||
dbClient,
|
||||
imageId: id,
|
||||
url,
|
||||
responseStatus: response.status,
|
||||
status: submitStatus,
|
||||
error: submitError,
|
||||
});
|
||||
|
||||
imageScanSubmittedCounter.inc({ lane: 'legacy', result: 'failed' });
|
||||
// createImageIngestionRequest's own log carries neither mediaType nor failureClass.
|
||||
logToAxiom({
|
||||
name: imageIngestionLogName(useImageScanning),
|
||||
type: 'error',
|
||||
reason: 'no-workflow-response',
|
||||
failureType: 'send-fail',
|
||||
failureClass,
|
||||
responseStatus: submitStatus,
|
||||
imageId: id,
|
||||
mediaType: type,
|
||||
}).catch(() => null);
|
||||
return false;
|
||||
}
|
||||
};
|
||||
|
||||
export const ingestImageBulk = async ({
|
||||
images,
|
||||
tx,
|
||||
lowPriority = true,
|
||||
scans,
|
||||
}: {
|
||||
images: IngestImageInput[];
|
||||
tx?: Prisma.TransactionClient;
|
||||
lowPriority?: boolean;
|
||||
scans?: ImageScanType[];
|
||||
}): Promise<boolean> => {
|
||||
if (!env.IMAGE_SCANNING_ENDPOINT)
|
||||
throw new Error('missing IMAGE_SCANNING_ENDPOINT environment variable');
|
||||
|
||||
const callbackUrl = env.IMAGE_SCANNING_CALLBACK;
|
||||
const scanRequestedAt = new Date();
|
||||
const imageIds = images.map(({ id }) => id);
|
||||
const dbClient = tx ?? dbWrite;
|
||||
|
||||
if (!imageIds.length) return false;
|
||||
|
||||
// TODO.articleImageScan: uncomment when ready to enable image scanning for articles
|
||||
// if (!isProd || !callbackUrl) {
|
||||
// console.log('skip ingest');
|
||||
// await dbClient.image.updateMany({
|
||||
// where: { id: { in: imageIds } },
|
||||
// data: {
|
||||
// scanRequestedAt,
|
||||
// scannedAt: scanRequestedAt,
|
||||
// ingestion: ImageIngestionStatus.Scanned,
|
||||
// nsfwLevel: NsfwLevel.PG,
|
||||
// },
|
||||
// });
|
||||
// return true;
|
||||
// }
|
||||
|
||||
const needsPrompts = !images.some((x) => x.prompt);
|
||||
if (needsPrompts) {
|
||||
const prompts = await dbClient.$queryRaw<{ id: number; prompt?: string }[]>`
|
||||
SELECT id, meta->>'prompt' as prompt FROM "Image" WHERE id IN (${Prisma.join(imageIds)})
|
||||
const scanJobsJson = JSON.stringify({ workflowId: workflowResponse.id });
|
||||
await dbClient.$executeRaw`
|
||||
UPDATE "Image"
|
||||
SET
|
||||
"scanRequestedAt" = ${scanRequestedAt},
|
||||
"scanJobs" = CASE
|
||||
WHEN "scanJobs" IS NOT NULL AND "scanJobs" ? 'retryCount' THEN
|
||||
${scanJobsJson}::jsonb || jsonb_build_object('retryCount', ("scanJobs"->'retryCount'))
|
||||
ELSE
|
||||
${scanJobsJson}::jsonb
|
||||
END
|
||||
WHERE id = ${id}
|
||||
`;
|
||||
const promptMap = Object.fromEntries(prompts.map((x) => [x.id, x.prompt]));
|
||||
for (const image of images) image.prompt = promptMap[image.id];
|
||||
}
|
||||
|
||||
const response = await fetch(
|
||||
env.IMAGE_SCANNING_ENDPOINT + `/enqueue-bulk?lowpri=${lowPriority}`,
|
||||
{
|
||||
method: 'POST',
|
||||
headers: { 'Content-Type': 'application/json' },
|
||||
signal: fetchTimeoutSignal(60_000),
|
||||
body: JSON.stringify(
|
||||
images.map((image) => ({
|
||||
imageId: image.id,
|
||||
imageKey: image.url,
|
||||
type: image.type,
|
||||
width: image.width,
|
||||
height: image.height,
|
||||
prompt: image.prompt,
|
||||
scans: scans ?? imageScanTypes,
|
||||
callbackUrl,
|
||||
}))
|
||||
),
|
||||
}
|
||||
);
|
||||
if (response.status === 202) {
|
||||
await dbClient.image.updateMany({
|
||||
where: { id: { in: imageIds } },
|
||||
data: { scanRequestedAt },
|
||||
});
|
||||
return true;
|
||||
}
|
||||
|
||||
return false;
|
||||
imageScanSubmittedCounter.inc({ lane, result: 'success' });
|
||||
return true;
|
||||
};
|
||||
|
||||
export function enqueueImageIngestion({
|
||||
@@ -6967,20 +6761,6 @@ export async function reportCsamImages({
|
||||
await bulkSetReportStatus({ ids: reportIds, status: ReportStatus.Actioned, userId: user.id, ip });
|
||||
}
|
||||
|
||||
export async function ingestArticleCoverImages(array: { imageId: number; articleId: number }[]) {
|
||||
const imageIds = array.map((x) => x.imageId);
|
||||
const images = await dbRead.image.findMany({
|
||||
where: { id: { in: imageIds } },
|
||||
select: { id: true, url: true, height: true, width: true },
|
||||
});
|
||||
|
||||
await articlesSearchIndex.queueUpdate(
|
||||
array.map((x) => ({ id: x.articleId, action: SearchIndexUpdateQueueAction.Update }))
|
||||
);
|
||||
|
||||
await ingestImageBulk({ images, lowPriority: true });
|
||||
}
|
||||
|
||||
export async function updateImageNsfwLevel({
|
||||
id,
|
||||
nsfwLevel,
|
||||
@@ -7282,15 +7062,16 @@ export async function addImageTools({
|
||||
select: {
|
||||
id: true,
|
||||
url: true,
|
||||
type: true,
|
||||
},
|
||||
});
|
||||
|
||||
if (updated.length > 0) {
|
||||
await ingestImageBulk({
|
||||
images: updated,
|
||||
lowPriority: true,
|
||||
});
|
||||
}
|
||||
enqueueImageIngestion({
|
||||
images: updated,
|
||||
name: 'add-image-tools',
|
||||
userId: user.id,
|
||||
lowPriority: true,
|
||||
});
|
||||
|
||||
for (const { imageId } of data) {
|
||||
purgeImageGenerationDataCache(imageId);
|
||||
|
||||
@@ -1,4 +1,5 @@
|
||||
import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';
|
||||
import type * as FliptClient from '~/server/flipt/client';
|
||||
|
||||
const {
|
||||
mockDbWrite,
|
||||
@@ -7,6 +8,7 @@ const {
|
||||
mockStringifyAIR,
|
||||
mockResolveDownloadUrl,
|
||||
mockIsProd,
|
||||
mockIsFlipt,
|
||||
} = vi.hoisted(() => ({
|
||||
mockDbWrite: {
|
||||
modelFile: { update: vi.fn() },
|
||||
@@ -19,6 +21,7 @@ const {
|
||||
// don't regress; failure-path tests reset to mockRejectedValue.
|
||||
mockResolveDownloadUrl: vi.fn().mockResolvedValue({ url: 'https://cdn.example/file' }),
|
||||
mockIsProd: { value: true },
|
||||
mockIsFlipt: vi.fn(),
|
||||
}));
|
||||
|
||||
vi.mock('~/server/db/client', () => ({ dbWrite: mockDbWrite }));
|
||||
@@ -78,6 +81,11 @@ vi.mock('~/env/server', () => ({
|
||||
},
|
||||
}));
|
||||
|
||||
vi.mock('~/server/flipt/client', async (importOriginal) => ({
|
||||
...(await importOriginal<typeof FliptClient>()),
|
||||
isFlipt: mockIsFlipt,
|
||||
}));
|
||||
|
||||
// orchestrator.service.ts pulls in edge-url which validates
|
||||
// NEXT_PUBLIC_* env vars at import-time. Stub it to keep tests hermetic.
|
||||
vi.mock('~/client-utils/edge-url', () => ({
|
||||
@@ -85,6 +93,7 @@ vi.mock('~/client-utils/edge-url', () => ({
|
||||
}));
|
||||
|
||||
import {
|
||||
createImageIngestionRequest,
|
||||
createModelFileScanRequest,
|
||||
ModelFileScanSubmissionError,
|
||||
} from '~/server/services/orchestrator/orchestrator.service';
|
||||
@@ -609,3 +618,130 @@ describe('createModelFileScanRequest', () => {
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
describe('createImageIngestionRequest', () => {
|
||||
const submittedSteps = () =>
|
||||
mockSubmitWorkflow.mock.calls[0][0].body.steps as Array<{ $type: string; input: unknown }>;
|
||||
|
||||
beforeEach(() => {
|
||||
mockIsFlipt.mockReset().mockResolvedValue(false);
|
||||
mockSubmitWorkflow.mockResolvedValue({
|
||||
data: { id: 'workflow-1' },
|
||||
response: { status: 200, headers: new Headers() },
|
||||
});
|
||||
});
|
||||
|
||||
it('submits wdTagging + mediaRating while the imageScanning flag is off', async () => {
|
||||
await createImageIngestionRequest({ imageId: 1, url: 'image-key' });
|
||||
expect(mockIsFlipt).toHaveBeenCalledWith('image-ingestion-image-scanning', '1');
|
||||
// The legacy submit is the production path; its inputs must stay exactly as they were.
|
||||
const metadata = { imageId: 1 };
|
||||
const priority = 'normal';
|
||||
const mediaUrl = { $ref: '$arguments', path: 'mediaUrl' };
|
||||
expect(submittedSteps()).toEqual([
|
||||
{
|
||||
$type: 'wdTagging',
|
||||
name: 'tags',
|
||||
metadata,
|
||||
priority,
|
||||
input: {
|
||||
mediaUrl,
|
||||
model:
|
||||
'urn:air:siglip2:repository:huggingface:cella110n/cl_tagger_v2@b57909b8e9c63f71e208a26473e7aabdf45ed6b6.tar',
|
||||
threshold: 0.55,
|
||||
},
|
||||
},
|
||||
{
|
||||
$type: 'mediaRating',
|
||||
name: 'rating',
|
||||
metadata,
|
||||
priority,
|
||||
input: {
|
||||
mediaUrl,
|
||||
engine: 'civitai',
|
||||
includeAgeClassification: true,
|
||||
includeAIRecognition: false,
|
||||
includeFaceRecognition: false,
|
||||
includeAnimeRecognition: false,
|
||||
},
|
||||
},
|
||||
{
|
||||
$type: 'mediaHash',
|
||||
name: 'hash',
|
||||
metadata,
|
||||
priority,
|
||||
input: { mediaUrl, hashTypes: ['perceptual'] },
|
||||
},
|
||||
]);
|
||||
});
|
||||
|
||||
it.each([
|
||||
[false, 'image-ingestion'],
|
||||
[true, 'image-scanning-ingestion'],
|
||||
])('logs a failed submit under its pipeline name (flag %s)', async (flagOn, name) => {
|
||||
vi.useRealTimers();
|
||||
mockIsFlipt.mockResolvedValue(flagOn);
|
||||
mockSubmitWorkflow.mockResolvedValue({
|
||||
data: undefined,
|
||||
error: 'bad request',
|
||||
response: { status: 400, headers: new Headers() },
|
||||
});
|
||||
const result = await createImageIngestionRequest({ imageId: 1, url: 'image-key' });
|
||||
|
||||
expect(result.useImageScanning).toBe(flagOn);
|
||||
expect(mockLogToAxiom).toHaveBeenCalledWith(expect.objectContaining({ name, imageId: 1 }));
|
||||
});
|
||||
|
||||
it('submits one imageScanning step when the flag is on', async () => {
|
||||
mockIsFlipt.mockResolvedValue(true);
|
||||
await createImageIngestionRequest({ imageId: 1, url: 'image-key' });
|
||||
const steps = submittedSteps();
|
||||
expect(steps.map((step) => step.$type)).toEqual(['imageScanning', 'mediaHash']);
|
||||
expect(steps[0].input).toEqual({ image: { $ref: '$arguments', path: 'mediaUrl' } });
|
||||
expect(steps[1].input).toEqual({
|
||||
mediaUrl: { $ref: '$arguments', path: 'mediaUrl' },
|
||||
hashTypes: ['perceptual'],
|
||||
});
|
||||
});
|
||||
|
||||
const frameUrl = { $ref: 'frame', path: 'url' };
|
||||
it.each([
|
||||
[
|
||||
false,
|
||||
[
|
||||
['wdTagging', 'mediaUrl'],
|
||||
['mediaRating', 'mediaUrl'],
|
||||
],
|
||||
],
|
||||
[true, [['imageScanning', 'image']]],
|
||||
])('repeats the per-frame steps over video frames (flag %s)', async (flagOn, templates) => {
|
||||
mockIsFlipt.mockResolvedValue(flagOn);
|
||||
await createImageIngestionRequest({ imageId: 1, url: 'video-key', type: 'video' });
|
||||
const steps = submittedSteps() as Array<{
|
||||
$type: string;
|
||||
input: {
|
||||
videoUrl?: unknown;
|
||||
for?: unknown;
|
||||
template?: { $type: string; input: Record<string, unknown> };
|
||||
};
|
||||
}>;
|
||||
|
||||
expect(steps[0]).toMatchObject({
|
||||
$type: 'videoFrameExtraction',
|
||||
input: {
|
||||
videoUrl: { $ref: '$arguments', path: 'mediaUrl' },
|
||||
frameRate: 1,
|
||||
uniqueThreshold: 0.9,
|
||||
maxFrames: 50,
|
||||
},
|
||||
});
|
||||
expect(steps.slice(1)).toHaveLength(templates.length);
|
||||
steps.slice(1).forEach((step, i) => {
|
||||
const [type, urlField] = templates[i];
|
||||
expect(step.$type).toBe('repeat');
|
||||
expect(step.input.for).toEqual({ $ref: 'videoFrames', path: 'output.frames', as: 'frame' });
|
||||
expect(step.input.template?.$type).toBe(type);
|
||||
expect(step.input.template?.input[urlField]).toEqual(frameUrl);
|
||||
});
|
||||
});
|
||||
});
|
||||
|
||||
@@ -18,6 +18,7 @@ import { logToAxiom } from '~/server/logging/client';
|
||||
import { internalOrchestratorClient } from '~/server/services/orchestrator/client';
|
||||
import { submitWorkflowWithRetry } from '~/server/services/orchestrator/workflows';
|
||||
import { hashContent } from '~/server/services/entity-moderation.service';
|
||||
import { FLIPT_FEATURE_FLAGS, isFlipt } from '~/server/flipt/client';
|
||||
import type { MediaType, ModelType } from '~/shared/utils/prisma/enums';
|
||||
import { EntityModerationStatus, ModelHashType, ScanResultCode } from '~/shared/utils/prisma/enums';
|
||||
import { stringifyAIR } from '~/shared/utils/air';
|
||||
@@ -47,6 +48,54 @@ const IMAGE_TAGGING_MODEL =
|
||||
'urn:air:siglip2:repository:huggingface:cella110n/cl_tagger_v2@b57909b8e9c63f71e208a26473e7aabdf45ed6b6.tar';
|
||||
const IMAGE_TAGGING_THRESHOLD = 0.55;
|
||||
|
||||
/** Axiom `name` for a failed ingestion submit, per scan pipeline. */
|
||||
export function imageIngestionLogName(useImageScanning: boolean) {
|
||||
return useImageScanning ? 'image-scanning-ingestion' : 'image-ingestion';
|
||||
}
|
||||
|
||||
type MediaUrlRef = { $ref: string; path: string };
|
||||
|
||||
function imageScanSteps({
|
||||
mediaUrl,
|
||||
metadata,
|
||||
priority,
|
||||
useImageScanning,
|
||||
}: {
|
||||
mediaUrl: MediaUrlRef;
|
||||
metadata: Record<string, unknown>;
|
||||
priority: Priority;
|
||||
useImageScanning: boolean;
|
||||
}) {
|
||||
if (useImageScanning)
|
||||
return [
|
||||
{ $type: 'imageScanning', name: 'scan', metadata, priority, input: { image: mediaUrl } },
|
||||
];
|
||||
|
||||
return [
|
||||
{
|
||||
$type: 'wdTagging',
|
||||
name: 'tags',
|
||||
metadata,
|
||||
priority,
|
||||
input: { mediaUrl, model: IMAGE_TAGGING_MODEL, threshold: IMAGE_TAGGING_THRESHOLD },
|
||||
},
|
||||
{
|
||||
$type: 'mediaRating',
|
||||
name: 'rating',
|
||||
metadata,
|
||||
priority,
|
||||
input: {
|
||||
mediaUrl,
|
||||
engine: 'civitai',
|
||||
includeAgeClassification: true,
|
||||
includeAIRecognition: false,
|
||||
includeFaceRecognition: false,
|
||||
includeAnimeRecognition: false,
|
||||
},
|
||||
},
|
||||
];
|
||||
}
|
||||
|
||||
export async function createImageIngestionRequest({
|
||||
imageId,
|
||||
url,
|
||||
@@ -68,6 +117,45 @@ export async function createImageIngestionRequest({
|
||||
// server-side, re-submitting with the same `externalId` returns the existing
|
||||
// workflow instead of duplicating it (orchestrator dedupes on (userId, externalId)).
|
||||
const externalId = randomUUID();
|
||||
const useImageScanning = await isFlipt(
|
||||
FLIPT_FEATURE_FLAGS.IMAGE_INGESTION_IMAGE_SCANNING,
|
||||
String(imageId)
|
||||
);
|
||||
const mediaUrl = { $ref: '$arguments', path: 'mediaUrl' };
|
||||
|
||||
const steps =
|
||||
type === 'image'
|
||||
? [
|
||||
...imageScanSteps({ mediaUrl, metadata, priority, useImageScanning }),
|
||||
{
|
||||
$type: 'mediaHash',
|
||||
name: 'hash',
|
||||
metadata,
|
||||
priority,
|
||||
input: { mediaUrl, hashTypes: ['perceptual'] },
|
||||
},
|
||||
]
|
||||
: [
|
||||
{
|
||||
$type: 'videoFrameExtraction',
|
||||
name: 'videoFrames',
|
||||
metadata,
|
||||
priority,
|
||||
input: { videoUrl: mediaUrl, frameRate: 1, uniqueThreshold: 0.9, maxFrames: 50 },
|
||||
},
|
||||
...imageScanSteps({
|
||||
mediaUrl: { $ref: 'frame', path: 'url' },
|
||||
metadata,
|
||||
priority,
|
||||
useImageScanning,
|
||||
}).map((template) => ({
|
||||
$type: 'repeat',
|
||||
input: {
|
||||
for: { $ref: 'videoFrames', path: 'output.frames', as: 'frame' },
|
||||
template,
|
||||
},
|
||||
})),
|
||||
];
|
||||
|
||||
const body: WorkflowTemplate = {
|
||||
externalId,
|
||||
@@ -76,110 +164,8 @@ export async function createImageIngestionRequest({
|
||||
mediaUrl: edgeUrl,
|
||||
},
|
||||
currencies: [],
|
||||
steps:
|
||||
type === 'image'
|
||||
? [
|
||||
{
|
||||
$type: 'wdTagging',
|
||||
name: 'tags',
|
||||
metadata,
|
||||
priority,
|
||||
input: {
|
||||
mediaUrl: { $ref: '$arguments', path: 'mediaUrl' },
|
||||
model: IMAGE_TAGGING_MODEL,
|
||||
threshold: IMAGE_TAGGING_THRESHOLD,
|
||||
},
|
||||
} as WorkflowStepTemplate,
|
||||
{
|
||||
$type: 'mediaRating',
|
||||
name: 'rating',
|
||||
metadata,
|
||||
priority,
|
||||
input: {
|
||||
mediaUrl: { $ref: '$arguments', path: 'mediaUrl' },
|
||||
engine: 'civitai',
|
||||
includeAgeClassification: true,
|
||||
includeAIRecognition: false,
|
||||
includeFaceRecognition: false,
|
||||
includeAnimeRecognition: false,
|
||||
},
|
||||
} as WorkflowStepTemplate,
|
||||
{
|
||||
$type: 'mediaHash',
|
||||
name: 'hash',
|
||||
metadata,
|
||||
priority,
|
||||
input: {
|
||||
mediaUrl: { $ref: '$arguments', path: 'mediaUrl' },
|
||||
hashTypes: ['perceptual'],
|
||||
},
|
||||
} as WorkflowStepTemplate,
|
||||
]
|
||||
: [
|
||||
{
|
||||
$type: 'videoFrameExtraction',
|
||||
name: 'videoFrames',
|
||||
metadata,
|
||||
priority,
|
||||
input: {
|
||||
videoUrl: { $ref: '$arguments', path: 'mediaUrl' },
|
||||
frameRate: 1,
|
||||
uniqueThreshold: 0.9,
|
||||
maxFrames: 50,
|
||||
},
|
||||
} as WorkflowStepTemplate,
|
||||
{
|
||||
$type: 'repeat',
|
||||
input: {
|
||||
for: {
|
||||
$ref: 'videoFrames',
|
||||
path: 'output.frames',
|
||||
as: 'frame',
|
||||
},
|
||||
template: {
|
||||
$type: 'wdTagging',
|
||||
name: 'tags',
|
||||
metadata,
|
||||
priority,
|
||||
input: {
|
||||
mediaUrl: {
|
||||
$ref: 'frame',
|
||||
path: 'url',
|
||||
},
|
||||
model: IMAGE_TAGGING_MODEL,
|
||||
threshold: IMAGE_TAGGING_THRESHOLD,
|
||||
},
|
||||
},
|
||||
},
|
||||
} as WorkflowStepTemplate,
|
||||
{
|
||||
$type: 'repeat',
|
||||
input: {
|
||||
for: {
|
||||
$ref: 'videoFrames',
|
||||
path: 'output.frames',
|
||||
as: 'frame',
|
||||
},
|
||||
template: {
|
||||
$type: 'mediaRating',
|
||||
name: 'rating',
|
||||
metadata,
|
||||
priority,
|
||||
input: {
|
||||
mediaUrl: {
|
||||
$ref: 'frame',
|
||||
path: 'url',
|
||||
},
|
||||
engine: 'civitai',
|
||||
includeAgeClassification: true,
|
||||
includeAIRecognition: false,
|
||||
includeFaceRecognition: false,
|
||||
includeAnimeRecognition: false,
|
||||
},
|
||||
},
|
||||
},
|
||||
} as WorkflowStepTemplate,
|
||||
],
|
||||
// WorkflowTemplate is @civitai/client's, which predates the imageScanning step.
|
||||
steps: steps as unknown as WorkflowStepTemplate[],
|
||||
callbacks: callbackUrl
|
||||
? [
|
||||
{
|
||||
@@ -229,7 +215,7 @@ export async function createImageIngestionRequest({
|
||||
if (!data) {
|
||||
logToAxiom({
|
||||
type: 'error',
|
||||
name: 'image-ingestion',
|
||||
name: imageIngestionLogName(useImageScanning),
|
||||
imageId,
|
||||
url,
|
||||
externalId,
|
||||
@@ -240,7 +226,7 @@ export async function createImageIngestionRequest({
|
||||
});
|
||||
}
|
||||
|
||||
return { data, body, error, status: response?.status };
|
||||
return { data, body, error, status: response?.status, useImageScanning };
|
||||
}
|
||||
|
||||
const PERCEPTUAL_HASH_WAIT_SECONDS = 30;
|
||||
|
||||
@@ -258,7 +258,7 @@ export async function auditPromptServer(options: AuditPromptOptions): Promise<vo
|
||||
// 🔴 Strip the NORMALIZED copy. `auditPromptEnriched` folds accents before the
|
||||
// detector runs, so stripping raw text matches one alphabet while the detector
|
||||
// reads another and a whitelisted `emma stone` still blocks `émma stone`. The
|
||||
// scan paths in image-scan-result.service.ts already normalize first.
|
||||
// scan paths in image-scan-pipeline.ts already normalize first.
|
||||
const [auditedPrompt, auditedNegativePrompt] = await Promise.all([
|
||||
stripBenignPhrases(normalizeText(prompt), BlocklistType.PromptBenignPhrase),
|
||||
stripBenignPhrases(normalizeText(negativePrompt), BlocklistType.NegativeBenignPhrase),
|
||||
|
||||
@@ -14,6 +14,7 @@
|
||||
import crypto from 'crypto';
|
||||
import type { MediaRatingOutput, Workflow, XGuardModerationStep } from '@civitai/client';
|
||||
import { clickhouse } from '~/server/clickhouse/client';
|
||||
import type { ImageScanningResult } from '~/server/services/image-scanning-result.service';
|
||||
import { logToAxiom } from '~/server/logging/client';
|
||||
import {
|
||||
applyDerivedLabels,
|
||||
@@ -530,3 +531,91 @@ export async function recordImageScan({
|
||||
labels,
|
||||
});
|
||||
}
|
||||
|
||||
// imageScanning's age and recognition models differ from mediaRating's, so its rows must not
|
||||
// aggregate with version 1.
|
||||
const IMAGE_SCANNING_AUDIT_VERSION = '2';
|
||||
|
||||
/** `recordImageScan` for the imageScanning step. */
|
||||
export async function recordImageScanningResult({
|
||||
workflowId,
|
||||
imageId,
|
||||
scan,
|
||||
startedAt,
|
||||
completedAt,
|
||||
}: {
|
||||
workflowId: string;
|
||||
imageId: number;
|
||||
scan: ImageScanningResult;
|
||||
startedAt?: Date | string | null;
|
||||
completedAt?: Date | string | null;
|
||||
}) {
|
||||
if (!clickhouse) return;
|
||||
|
||||
const baseRow = {
|
||||
threshold: null,
|
||||
version: IMAGE_SCANNING_AUDIT_VERSION,
|
||||
matchedText: [] as string[],
|
||||
matchedPositivePrompt: [] as string[],
|
||||
matchedNegativePrompt: [] as string[],
|
||||
};
|
||||
const labels: LabelRowSeed[] = [
|
||||
{
|
||||
...baseRow,
|
||||
label: normClassifierLabel(scan.nsfwLevel),
|
||||
labelValue: 'nsfw_level',
|
||||
score: 1,
|
||||
triggered: 1,
|
||||
},
|
||||
];
|
||||
|
||||
// Nothing acts on this label yet.
|
||||
if (scan.csam !== null) {
|
||||
labels.push({
|
||||
...baseRow,
|
||||
label: 'csam',
|
||||
labelValue: '',
|
||||
score: scan.csam ? 1 : 0,
|
||||
triggered: scan.csam ? 1 : 0,
|
||||
});
|
||||
}
|
||||
|
||||
if (scan.ageDetections.length > 0) {
|
||||
labels.push({
|
||||
...baseRow,
|
||||
label: 'minor',
|
||||
labelValue: scan.ageDetections
|
||||
.map((d) => d.ageBand)
|
||||
.filter((x): x is string => !!x)
|
||||
.join(', '),
|
||||
score: clamp01(Math.max(...scan.ageDetections.map((d) => d.under18Probability ?? 0))),
|
||||
triggered: scan.minorDetected ? 1 : 0,
|
||||
});
|
||||
}
|
||||
|
||||
for (const [recognition, labelValue] of [
|
||||
[scan.aiRecognition, 'ai_recognition'],
|
||||
[scan.animeRecognition, 'anime_recognition'],
|
||||
] as const) {
|
||||
if (!recognition?.label) continue;
|
||||
labels.push({
|
||||
...baseRow,
|
||||
label: normClassifierLabel(recognition.label),
|
||||
labelValue,
|
||||
score: clamp01(recognition.score),
|
||||
triggered: 1,
|
||||
});
|
||||
}
|
||||
|
||||
await insertRows({
|
||||
workflowId,
|
||||
scanner: 'image_ingestion',
|
||||
contentHash: computeContentHash(`image:${imageId}`),
|
||||
entityType: 'image',
|
||||
entityId: String(imageId),
|
||||
modelVersion: IMAGE_SCANNING_AUDIT_VERSION,
|
||||
startedAt,
|
||||
completedAt,
|
||||
labels,
|
||||
});
|
||||
}
|
||||
|
||||
@@ -94,9 +94,8 @@ export async function debounceArticleUpdate(articleId: number): Promise<void> {
|
||||
* advance article state, and `updateArticleNsfwLevels` picks up the new cover
|
||||
* rating so `Article.nsfwLevel` no longer lags the cover.
|
||||
*
|
||||
* Callers are responsible for gating on the `articleImageScanning` feature flag
|
||||
* because the two webhook entrypoints resolve the flag differently (one reads
|
||||
* from a request context, the other receives it as a parameter).
|
||||
* Callers gate on the `articleImageScanning` feature flag — it needs a request context,
|
||||
* which this helper does not have.
|
||||
*/
|
||||
export async function fanOutArticleImageUpdates(imageId: number): Promise<void> {
|
||||
const [articleConnections, coverArticles] = await Promise.all([
|
||||
|
||||
@@ -21,9 +21,8 @@ import { trpcMutation, trpcQuery, uniqueToken } from './preview-trpc';
|
||||
* bucket (a tiny PNG is written each run — harmless).
|
||||
*
|
||||
* Scope ceiling (honest): asserts the upload + attach + PUBLISH DB path. It does NOT
|
||||
* assert public-feed visibility — image ingestion/scan is unreachable in preview
|
||||
* (placeholder IMAGE_SCANNING_ENDPOINT), so the row stays ingestion:Pending. That does
|
||||
* NOT block row creation, attach, or publish.
|
||||
* assert public-feed visibility — image ingestion never completes in preview, so the row
|
||||
* stays ingestion:Pending. That does NOT block row creation, attach, or publish.
|
||||
*
|
||||
* Role: tester (free member, passes the gate; guardedProcedure cleared via onboarding=15).
|
||||
*/
|
||||
|
||||
Reference in New Issue
Block a user