The throughput bottleneck was curator-side, not the network. lease() claimed the
lowest-id pending/expired jobs with `... ORDER BY id LIMIT n`, but with only a
plain `status` index Postgres walked the primary key from id=1, skipping the
entire prefix of already done/error rows before reaching pending ones. As `done`
grew (69k+), every lease became an O(done) scan — leasing crawled, the DB
saturated, and even /status (the queue GROUP BY count) stalled the agent.
- Migration 0070 adds two partial indexes over just the live slice: pending rows
indexed by id (hot path), and leased rows by lease_expires_at (crash-recovery
+ orphan sweep). They stay tiny no matter how large the done/error history.
- lease() split into two phases so each uses a partial index: claim pending
first (id-ordered, O(batch)); reclaim expired leases only when pending can't
fill the batch. Same semantics (SKIP LOCKED, attempts++, expired reclaim).
- Model __table_args__ declares the indexes so ORM and schema agree.
- Test: a done-prefix at low ids must not stop the lease reaching pending.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Heads + CCIP are the tag source and head auto-apply is the earned propagation.
The Camie tagger ran only to feed the allowlist bulk-apply (its ImagePrediction
rows had no other consumer), and the allowlist was a SECOND, un-earned auto-apply
path firing in parallel with heads on every accept — exactly the un-earned spray
the v2 pivot replaced. Retire both.
Behavior change: accepting a suggestion now applies the tag to THAT image only
(source='ml_accepted', a head-training positive) — it no longer allowlists +
fans the tag across the library via Camie. Propagation is heads' earned
auto-apply. (Loses instant cold-start propagation for booru-vocab tags; that was
un-earned and bypassed the precision gate.)
- tag_and_embed is now EMBED-ONLY (no Camie load/infer, no ImagePrediction
writes); backfill enqueues it for images with no embedding.
- Removed: services/ml/tagger.py, apply_allowlist_tags + helpers + daily beat +
every enqueue caller (accept/alias/merge/per-image), api/allowlist.py +
blueprint, ImagePrediction + TagAllowlist models/tables (migration 0067),
AllowlistTable.vue + allowlist store, the accept coverage-projection payload.
- AllowlistService gutted to accept/dismiss/undismiss/reject (the rejection store
the rail still needs); accept returns nothing, API returns {accepted, tag_id}.
- tag merge no longer repoints/triggers the allowlist; _keep_as_alias now keys on
ML-applied image_tag sources (incl. head_auto) instead of the allowlist.
- UI: MLBackfillCard relabelled to embedding-only; accept toast simplified;
MaintenancePanel drops the allowlist tile.
Left for a follow-up hygiene pass (now-inert, harmless): the dead settings
columns (tagger_store_floor, tagger_model_version, suggestion_threshold_*,
video_min_tag_frames), image_record.tagger_model_version, MLThresholdSliders
trim, and the Camie model download in download_models.py.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
The v2 pivot replaced per-tag SigLIP centroids with learned heads + CCIP.
Centroids were still recomputed (on every tag merge + a daily beat) but NOTHING
read them — suggestions come from heads+CCIP and apply_allowlist_tags applies
via Camie predictions, not centroids. Pure dead wiring; remove it.
Removed: CentroidService, recompute_centroid/recompute_centroids tasks, the
daily beat, POST /api/ml/recompute-centroids, the recompute-on-merge trigger,
the tag_reference_embedding table + model, the centroid_similarity_threshold +
min_reference_images settings (migration 0066), the CentroidRecomputeCard +
its store action + MaintenancePanel tile, and the centroid slider in
MLThresholdSliders. _keep_as_alias drops its vestigial has-centroid branch (the
allowlist branch already covers "could re-emit"); tag merge no longer clears a
table that no longer exists.
NOT touched (still live, parallel to heads): the Camie tagger, ImagePrediction,
and the allowlist bulk-apply — accepting a suggestion still allowlists + applies
it across the library. The tag-eval "centroid" baseline metric is unrelated
(in-memory) and stays. (image_record.centroid_scores JSON column also remains —
separate legacy field, its own micro-cleanup.)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Make the SigLIP embedder an operator choice (drop-in to SigLIP 2:
google/siglip2-so400m-patch16-512 is a verified 1152-d model at 512px → no
schema change, better small-cue fidelity). A swap = set model + re-embed +
retrain, all operator-driven; the GPU agent does the re-embed so it's fast.
- settings: embedder_model_name is now a setting (migration 0065) alongside the
existing embedder_model_version; both editable + validated (non-empty) in the
ml admin API. The server embedder loads by HF name (AutoImageProcessor/Model,
model-agnostic), preferring the pre-downloaded local dir for the default so
existing deploys don't re-download; rebuilds on a name change.
- agent: new 'embed' job = whole-image SigLIP embedding (mean-pool video frames)
under the lease-announced model → POST /jobs/submit_embedding writes
image_record.siglip_embedding + siglip_model_version. The lease now announces
the model FROM THE SETTING (not a constant).
- re-embed routing: enqueue_gpu_backfill('embed') selects unembedded + stale-
version images; 'siglip' now re-embeds concept crops whose version != current
(so a swap re-triggers crops, not just the never-embedded back-catalogue). The
CPU ml-worker backfill no longer re-embeds on a version mismatch (it can't
churn the library at 512px) — the GPU agent owns version re-embeds. Daily
'embed' + 'siglip' beats self-heal.
- scoring: score_image only bags embeddings in the CURRENT model's space (whole-
image gated by siglip_model_version, concept regions by embedding_version) so a
mid-swap stale vector isn't scored by new-space heads; legacy NULL = current.
- UI: GpuAgentCard "Embedding model (advanced)" — edit name/version, Save, and
"Re-embed library (GPU)" (queues embed + siglip); points at SigLIP 2.
Tests: lease announces model + submit_embedding round-trip; enqueue 'embed'
selects stale/unembedded; stale-version excluded from scoring; embedder model
settable + empty rejected; siglip gate updated to current-version concept.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Lift recall on small/local concepts (glasses, cum, stomach-bulge, xray,
lactation) that the whole-image SigLIP vector washes out: the GPU agent now
embeds figure crops with SigLIP too, stored as kind='concept' regions, and the
suggestion rail scores each image as a BAG (whole-image + every concept crop),
taking each head's MAX over the bag. The whole-image vector is always in the
bag, so this can never score lower than before.
Model-agnostic by construction: the server ANNOUNCES the embedding model
(HF name + version) in the lease, so the agent loads whatever the heads were
trained in and stays in lock-step — a model swap is a server setting + a
re-embed migration, never an agent change.
- agent: model-agnostic CropEmbedder (torch/transformers get_image_features,
fp16 on CUDA, inference-locked); worker branches on job.task — 'ccip' emits
figure(CCIP)+concept(SigLIP) in one pass, 'siglip' emits concept-only so the
back-catalogue backfill never churns figure/CCIP regions; torch cu124 +
transformers in the image.
- server: lease announces embed_model_name/embed_version; score_image is
max-over-bag (version-filtered region embeddings); enqueue_gpu_backfill
'siglip' gates on a missing concept region (drains the back-catalogue,
retries failures, no double-enqueue); daily siglip-backfill beat; UI button;
/api/ccip/overview reports images_with_concept_siglip.
- v1 scope: suggestion rail only — auto-apply stays whole-image (conservative;
heads' thresholds were calibrated on whole-image). Bulk-apply bag = follow-up.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Works through the optional CCIP ideas + the "keep moving even if I forget" ask:
AUTOMATION (no button needed):
- Hourly beat auto-enqueues CCIP backfill — new images get embedded (and errored
ones retried) on their own; the queue never goes idle waiting for a click.
- CCIP auto-apply: a daily sweep tags confident matches (source='ccip_auto') so
identity tags keep flowing. ON by default (opt-out, like head auto-apply);
ml_settings.ccip_auto_apply_enabled + _threshold (0.92, above the suggest cut),
migration 0064. Vectorized (one matmul + reduceat per image), reversible, skips
already-applied/rejected. Switch + threshold in the GPU agent card; GET/PATCH
/api/ml/settings; auto_applied count in /api/ccip/overview.
REFERENCE QUALITY (the over-fire root cause):
- character_references now draws ONLY from single-character images — on a
multi-character image the tag is image-level, so every figure would otherwise
pollute each character's prototypes (a 2-char image tagged 'Velma' made
Daphne's figure a Velma reference). This is the contamination behind residual
over-firing.
- Cached on a cheap signature (char-tag count + ccip-region count/max-id) so the
reference load isn't redone on every modal open.
Tests: multi-character image not used as a reference; auto-apply tags a confident
match as ccip_auto.
NEXT (not done, confirmed): comic-panel cropping + SigLIP concept crops ("spot
interesting content").
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Live data showed the v1 flat 0.75 cosine over-fired — ~64% of matched images got
3-10 character guesses dominated by the most-referenced characters (a 27-ref
character clears a low bar on many images). A sweep showed 0.85 collapses the
noise (noisy multi-matches 47→3) while keeping the confident single-character
matches.
- ml_settings.ccip_match_threshold (migration 0063, default 0.85); match_image
reads it (override still accepted). DEFAULT_SIM_THRESHOLD fallback 0.75→0.85.
- Exposed in GET/PATCH /api/ml/settings (validated 0.5–0.999).
- Slider in the GPU agent card ("Character-match strictness") — tune live, no
redeploy, same observe-and-tune loop as auto-apply.
Test: a ~0.9-cosine figure matches at 0.85, dropped at 0.95.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
So work an agent orphaned gets picked back up quickly, three layers:
- GpuJobService.release(): a graceful agent stop hands its still-leased jobs back
to pending instantly (POST /api/gpu/jobs/release), no waiting out the lease.
- GpuJobService.recover_orphaned() + recover_orphaned_gpu_jobs Celery task on a
60s beat: resets expired leases (a hard-crashed agent) to pending and keeps the
queue counts honest even when nothing is leasing.
- Lease TTL 300→180s: still well above any single job (a capped-frame video embed
is tens of seconds, and a live worker heartbeats), but a hard crash recovers
faster once the sweep fires.
Tests: release returns-to-pending (token-scoped), recover_orphaned resets only
expired leases, release API round-trip.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
SuggestionService.for_image now merges CCIP character matches with the SigLIP
head suggestions — they're complementary, not exclusive: CCIP is the identity-
specialized signal but needs a detected figure; the heads work whole-image but
conflate identity with style. Merged by tag: 'both' when they corroborate
(higher score wins), 'ccip' / 'head' otherwise. Cheap when no CCIP vectors exist
yet (match_image returns early without a figure vector), so it's a no-op until
the agent runs. Suggestion.source is now 'head' | 'ccip' | 'both'.
Test: a character with a CCIP reference figure surfaces (source='ccip') on a new
image whose figure matches.
NEXT: the agent container (real CCIP/detector models, hands-on) that produces the
vectors this consumes.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
The server-side brain that turns stored CCIP vectors into character suggestions
— no GPU. character_references() gathers each character tag's prototype vectors
(figure/face-region CCIP embeddings on images carrying that tag); match_image()
cosine-matches an image's figure vectors against every character (multi-
prototype: best over a character's examples), surfacing those above a tunable
threshold as {tag_id, name, category:'character', score, source:'ccip'},
excluding already-applied characters. v1 = cosine on raw CCIP vectors; the exact
CCIP metric/threshold gets validated against the model in the hands-on eval.
Tests (synthetic vectors): same-character match across images, no-match for an
orthogonal figure, already-applied exclusion, no-figure-vectors empty.
NEXT: merge CCIP character suggestions into the rail; the agent container that
actually produces the vectors (hands-on, GPU — not CI-verifiable).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Answers "how are videos/all media handled by the GPU worker": a job is per ITEM,
but the agent fans a VIDEO into per-frame instances (ffmpeg in the agent, the
existing cadence), each stored with a timestamp — so a video becomes a BAG of
frame embeddings (fixes the mean-embedding muddle) instead of one washed-out
vector. Stills → frame_time NULL; animated GIF/WebP treated like short video.
- image_region.frame_time (migration 0061, not yet deployed so folded in): the
source frame's seconds for video/animated media; NULL for stills. RegionService
passes it through. A whole frame is just kind='frame'.
- gpu_job + GpuJobService (migration 0062): the durable work list that keeps the
desktop agent HTTP-only — enqueue (dedupes (image,task)) / lease (FOR UPDATE
SKIP LOCKED, re-claims expired leases so the queue self-heals) / heartbeat /
complete / fail (re-queues until MAX_ATTEMPTS then 'error'). The server enqueues;
the agent leases+submits over the web API; Redis/Postgres stay private.
Tests: enqueue dedupe, lease-then-skip-when-held, expired-lease reclaim, scoped
heartbeat, complete, fail-requeue-then-error. region test now covers frame_time.
NEXT: the thin HTTP API (lease/submit/heartbeat) + bearer-token auth, then the
agent container + control UI.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
The storage backbone both crop jobs write to and read from. image_region =
normalized bbox (rx/ry/rw/rh) + kind ('face'/'figure' → CCIP character id;
'concept' → SigLIP head bag) + the crop's embedding (nullable Vector(768) CCIP /
Vector(1152) SigLIP, one per kind) + version stamps for compute-once gating. The
bbox doubles as grounded-tag provenance. Migration 0061.
RegionService.replace_regions (scoped BY KIND so the figure + concept pipelines
don't clobber each other) + get_regions — the GPU agent's results endpoint will
call the writer; the character matcher + bag scorer read. Server-side, no GPU.
Tests: replace/get round-trip, kind-scoped replacement, CCIP vector round-trip.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
The trunk of both crop jobs — CCIP figure-crops and SigLIP concept-crops call
the SAME crop_region(): normalized-bbox crop with optional context padding,
edge-clamping, and the lower-bound size floor (max of a fraction-of-short-side
and an absolute pixel floor) below which a region is too small to embed and
returns None. Only the proposer (where) and embedder (what) differ; the crop is
shared. Pure Pillow — importable + testable anywhere (the GPU agent imports it
for the crop step). Unit-lane tests (no DB): region pixels, floor rejection,
edge clamp, pad expansion, out-size resize.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Graduated heads can now apply their tag without a human — gated so it's safe:
- FIRING GATE: a head fires only when the master switch (head_auto_apply_enabled,
default OFF) is on AND it has >= head_auto_apply_min_positives (default 30)
clean labels. A precise-looking but under-supported low-N head can't spray tags.
- auto_apply_sweep (heads.py): streams every embedded image in chunks, scores
against the eligible heads (numpy, no sklearn), applies each head's tag where
score >= its auto_apply_threshold and the tag isn't already applied/rejected,
with source='head_auto' (distinguishable + reversible). dry_run counts only.
- HeadAutoApplyRun (migration 0059) tracks each sweep / preview; apply_head_tags
task (ml queue) + scheduled_apply_head_tags daily beat (no-op unless enabled)
+ recovery sweep + retention(20).
- API: POST /api/heads/auto-apply {dry_run} (202 / 409 running / 400 disabled),
GET /api/heads/auto-apply (recent runs + per-concept report). Settings
head_auto_apply_enabled + min_positives via /api/ml/settings.
Tests: sweep applies above threshold, dry-run writes nothing, skips under-
supported + ungraduated heads; API disabled/dry-run/conflict guards.
NEXT (slice 2): the observability the operator asked for — per-concept misfire
(auto-applied-then-removed) + under-fire tracking, time-series snapshots, and a
reporting API to tune. Slice 3: the UI (enable, preview, trends).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
The rail's Suggestions now come from the trained per-concept heads. SuggestionService.for_image scores the image's frozen SigLIP embedding against
every head (heads.score_image) and surfaces concepts above each head's own
suggest threshold; the typed-dropdown's min=0 "show everything" mode maps to a
flat floor so any head-scored concept can still be picked. Already-applied tags
drop; rejected tags stay flagged + reversible (unchanged).
REMOVED from the suggestion path (rule 22, no fallback): the Camie
ImagePrediction candidate/alias/merge pipeline and the per-tag centroid
augmentation, plus the now-dead SuggestionService internals (_load_predictions,
_threshold_for, _settings, self.aliases, self.centroids). Head suggestions are
always canonical tags, so raw_name/via_alias are null/false and the rail's
alias kebab is inert by data (its removal + the Camie ingest-tagger rip are the
flagged follow-up). for_selection (bulk consensus) now aggregates head
suggestions unchanged.
Tests rewritten to the head path: test_ml_suggestions (surfaces/applied/
rejected-reversible/override/no-embedding/no-heads), test_suggestions_bulk
(consensus), test_api_suggestions (get + dropped the Camie-alias roundtrip),
and test_ml_artist_retired (artist not head-eligible via _HEAD_KINDS).
DEPLOY NOTE: after this lands, the rail is empty until you run Train heads
(Settings → Tagging → Concept heads) — deploy, train, then the rail populates.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
Alphabetize HeadTrainingRun in models/__init__ + maintenance imports (H before
I), and drop the inline comment that split heads.py's import block. Pure import
ordering — no behavior change. (run 1601 lint)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
The eval (#1130) proved the frozen-embedding + trained-head spine; this lands
its production form (the first of three slices that make heads the suggestion
source, replacing Camie + centroid).
- tag_head: one logistic-regression head per general/character concept with
enough labelled positives. Weights (pgvector), honest CV-derived suggest
threshold + earned-auto-apply point, and per-concept quality metrics.
- head_training_run: persisted batch lifecycle (mirrors tag_eval_run) so the
admin card shows live + historical status across navigation.
- services/ml/heads.py: TRAIN (sync, ml worker, reuses tag_eval's proven data
loaders + metric math so production heads match measured eval numbers) and
SCORE (async, API worker — numpy via pgvector, no scikit-learn): score one
image's embedding against all heads → the rail's suggestions, cached on
(count, max trained_at) so a retrain invalidates without per-request loads.
- tasks.ml.train_heads (ml queue, commits per head so a kill leaves progress)
+ recover_stalled_head_training_runs sweep + retention(20) + 5-min beat
(rule 89).
- api/heads.py: POST /api/heads/train (one run at a time, 409 guard) + GET
/api/heads (count, graduated, last-trained, running, per-concept table,
recent runs).
- ml_settings: head_min_positives + head_auto_apply_precision, tunable via
/api/ml/settings.
Scoring isn't wired into the rail yet (slice C) and the admin UI is slice B —
this slice makes training + scoring exist and CI-verifiable. 'precision' column
stored as precision_cv (SQL reserved word). Migration 0058.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
A red-✗ dismissal no longer makes the suggestion vanish. The rejected tag
stays in the rail — dimmed, struck-through, with a "rejected" pill and a
one-click undo (↶) in place of the ✗ — so a misclick is recoverable and the
operator can see what they've said no to (operator-asked 2026-06-27).
Backend: SuggestionService.for_image now KEEPS rejected tags, flagged
rejected=True, sorted to the bottom of their category, instead of dropping
them. New AllowlistService.undismiss + POST /suggestions/undismiss clears the
TagSuggestionRejection. Rejected items are still excluded from bulk consensus
(for_selection) and the type-to-add dropdown, whose jobs are unchanged.
Frontend: store.dismiss flags in place (canonical tags) rather than dropping;
new store.undismiss reverts. SuggestionItem renders the rejected state and
swaps ✗→↶; ✓ still accepts (which clears the rejection server-side).
Tests: rejected-surfaced-flagged-then-reversible (service) + undismiss
endpoint idempotency (API).
Completes #1134's reversible-rejection half. Heads-as-suggestion-source is
the remaining piece.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ttrj5P7upUTueSfoJcxEqa
"Keep" on a doubted positive was a no-op, so the same confirmed-correct images
came back in "head doubts" every run (operator-flagged: reinforcement keeps
surfacing the same images). Add tag_positive_confirmation (mirror of
tag_suggestion_rejection): keep → POST /images/<id>/tags/<tag_id>/confirm, and
the eval excludes confirmed positives from the doubts list — exactly as rejected
items already drop out of the suggest list. The tag stays a positive either way
(confirmation is a "reviewed" marker, not a training change).
- model TagPositiveConfirmation + migration 0057; confirm endpoint (idempotent).
- tag_eval: _confirmed_ids + exclude from head_doubts_positive examples.
- store.confirmTag + card "keep" calls it.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
"head would suggest" drew from the whole negative pool, which INCLUDES the
images the operator rejected. A rejected near-miss (e.g. an orc under "goblin")
is a hard negative that still scores high, so it kept resurfacing as a fresh
suggestion every run (operator-flagged: "same items keep appearing"). Exclude
already-rejected ids from the suggest list — once you've said no, it's gone.
(head doubts = lowest-scoring positives is unchanged; genuinely-hard true
positives legitimately recur there.)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Two additions driven by "what's the commit threshold?" + "find more tags":
1. High-precision operating point (Bar 4). Per concept, report the threshold that
maximizes recall while holding precision >= a target (default 0.97, configurable
via `precision_target`) — i.e. "could this fire without a human, and how much
would it catch?" `head.auto_apply` = {target, threshold, precision, recall} or
null if the target is unreachable. Surfaced on the card.
2. Server-side concept auto-discovery. `auto_top_n` param unions the explicit
concept list with the N most-tagged general tags (one fast DB query) so the
eval can broaden itself without hand-listing — replaces the slow HTTP directory
paging. Card gains "+ auto-add top-N" and precision-target inputs.
No migration; numpy/sklearn stay lazy. Existing _normalize_params test still
holds (new keys additive; None still falls back to DEFAULT_CONCEPTS).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Frontend for #1130. A maintenance tile in Settings → Tagging:
- Editable concept list + "Run eval" → POST /api/tag-eval (one running at a time).
- Rehydrates on mount via the persisted run (getRun by latest id) and polls while
running — so the report SURVIVES navigation (operator-flagged); the task runs
backend-side regardless and the card reconnects to its row.
- Renders the saved report: per-concept head-vs-centroid metrics table (AP/F1/
precision/recall) with Δ AP, the learning curve (AP @ N positives), and
thumbnail galleries (head-would-suggest / head-doubts-positive) for eyeballing.
Backend: _examples now stores thumbnail_urls (not just ids) so the report is a
self-contained artifact that renders without per-id lookups on reload.
No new top-level surface — slots into the existing maintenance area.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Slice 1 of milestone #114 (tagging v2). Proves the frozen-embedding + trained-
head spine on the operator's own data, reusing the SigLIP embeddings already
stored on image_record — no re-embedding, no GPU.
Per concept: train a logistic-regression HEAD (positives + negatives = explicit
rejections + sampled unlabeled) vs the old single-CENTROID baseline; report
cross-validated precision/recall/AP for both, a LEARNING CURVE (AP/F1 as tagged
positives grow 10→30→100→300), and example image ids (head-would-suggest /
head-doubts-positive) to eyeball.
Persisted so the report SURVIVES navigation (operator-flagged): the run + full
report live in a new tag_eval_run row (mirrors library_audit_run); the admin
card will rehydrate from GET on mount, not transient state.
- models.TagEvalRun + migration 0056; runs on the ml queue (only worker with
numpy/sklearn) — numpy/sklearn lazy-imported so the API can still enqueue.
- services/ml/tag_eval (compute + start helper, one-running guard), tasks.ml
.tag_eval_run, api/tag-eval (POST create, GET history light / detail w/ report).
- recover_stalled_tag_eval_runs sweep + retention (keep last 20) + 5-min beat
(rule 89). scikit-learn added to requirements-ml.
- tests: param normalization + the rehydrate read-path + create/conflict.
Frontend admin card (trigger + render persisted report) follows next.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Cluster B, milestone #99. Backend for the allowlist tuning dashboard.
#7a: AllowlistService.coverage(tag_id, threshold) counts distinct images with
a prediction resolving to the tag (raw_name==tag.name OR (raw_name,category) in
the tag's aliases) scoring >= threshold — the gross candidate pool, mirroring
tasks.ml._confidence_for_tag resolution. list_all now carries applied_count
(grouped image_tag count) + coverage_count (at the row's threshold). New
GET /api/tags/<id>/allowlist/coverage?threshold= for the live what-if number.
#7b: /suggestions/accept + /alias return {allowlisted, tag_id, tag_name,
projected_count} (projection at the tag's threshold) instead of 204, so the UI
can show a non-blocking 'auto-applying to ~N images' toast. Apply still runs
async via apply_allowlist_tags — projected_count is an estimate.
Tests: coverage by threshold (direct + alias-with-category), list applied vs
coverage, coverage route (explicit/default/bad threshold), accept/alias payload
(newly-allowlisted vs already-on-list).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XCUHUGQLrBrkgyk1t49kpX
GPU enablement (#872) cancelled — not worth the Pascal-specific build for a
modest CPU→GPU win on an old P4. Remove the dead GPU code (device.py, the CUDA
provider branch in tagger, the .to('cuda') path in embedder) so nothing carries
it forward.
Instead, bound CPU inference threads by default so the ml-worker is a predictable
core consumer on a SHARED node — the intended scaling model is multiple worker
replicas (each --concurrency=1, each its own cgroup limit), not one big
container. ONNX Runtime and torch otherwise size their thread pools to ALL host
cores, so each replica would grab every core and oversubscribe / starve the
co-located DB+web. Cap both to _INTRA_OP_THREADS=4 (matches the prior per-worker
cpus:4 unit): run N replicas where N×4 stays within the cores allotted to ML.
- tagger: ort.SessionOptions().intra_op_num_threads = 4 (CPUExecutionProvider).
- embedder: torch.set_num_threads(4).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Step 1 of GPU enablement (code only — CPU-safe, CI-green; the CUDA image is a
separate step pending the host driver version).
- New services/ml/device.py: FC_ML_DEVICE (auto|cuda|cpu) intent + VRAM knobs
(FC_ML_ONNX_GPU_MEM_GB, FC_ML_TORCH_MEM_FRACTION). Per-worker-host bootstrap →
env, not a DB setting (the GPU host runs CUDA, others CPU).
- tagger: use CUDAExecutionProvider (with gpu_mem_limit) when requested AND the
provider is actually present (onnxruntime-gpu), else CPUExecutionProvider. Logs
the active providers.
- embedder: move model + inputs to cuda when requested AND torch.cuda is
available; cap torch's VRAM share; .detach().cpu() before numpy. fp32 kept so
GPU embeddings stay in the same space as existing CPU ones.
Both AND the env intent with the framework's real availability, so on CPU
(CI / CPU onnxruntime / no GPU) they fall back cleanly — behavior unchanged.
The 8GB P4 is shared by both frameworks, hence the conservative default caps.
Tests: device env parsing. (tagger/embedder GPU paths are operator-verified on
the GPU host — models aren't in CI.)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The headline bug: aliases created from the modal NEVER resolved. Create
sent the normalized display name ('Sword', 'Uchiha Sasuke') while
resolution keys on the raw booru model key ('sword', 'uchiha_sasuke',
case-sensitive) — so the mapping was stored under a key nothing looks up,
and the prediction kept reappearing unaliased. The raw key wasn't even in
the /suggestions response, so the modal couldn't send it.
- Suggestion now carries raw_name (the model key an alias must use) and
via_alias (surfaced via an operator alias); both serialized by the API.
- Modal alias-create sends raw_name, not display_name (the fix). Aliased
suggestions show an 'alias' badge and a 'Remove alias' action; 'Treat as
alias for…' is hidden for centroid hits (no model key) and already-aliased
rows.
- Tag-side management: TagCard ⋮ → 'Aliases…' opens a dialog listing the
model keys that fold into a tag, with remove (GET /api/tags/<id>/aliases +
AliasService.list_for_tag). Creation stays in the modal suggestion flow.
Tests: full API round-trip locking the raw-key contract (raw_name exposed →
alias authored with it → resolves + via_alias on a later image);
list_for_tag (service + API); via_alias/raw_name on the existing service
suggestion tests. No migration.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Read cutover verified in prod (suggestions + allowlist read image_prediction;
backfill complete at 908k rows / 51k images). Removes the old JSON column and
everything that fed it:
- ImageRecord.tagger_predictions column removed; migration 0046 DROPs it.
tagger_model_version kept as the "tagged / current?" signal the backfill
sweep reads (needs-tagging check switched to tagger_model_version IS NULL).
- tag_and_embed no longer dual-writes the JSON — image_prediction is the only
write path.
- importer re-import reset drops the JSON line (image_prediction rows are
already deleted on re-import).
- Retired the one-time #768 backfill task + the #764 prune task, their admin
endpoints, and their Maintenance cards (Backfill/PrunePredictionsCard).
- Tests seed/assert via image_prediction; stale column refs removed.
Disk reclaim is NOT automatic: DROP COLUMN is a catalog change. Run
`VACUUM FULL image_record` off-hours afterward to return the ~100 GB to the OS
so DB backups go small (#739). image_prediction (~90 MB) stays in pg_dump — it's
the source of truth now.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Switch every prediction READER off the JSON column onto the normalized
image_prediction table. Parity by construction: each reader loads the same
{raw_name: {category, confidence}} dict it consumed before (via small
_load_predictions helpers), so all downstream threshold/alias/merge/consensus
logic is byte-identical — only the data source changed.
- suggestions.SuggestionService.for_image (and for_selection via it)
- ml.apply_allowlist_tags (iterates images that have prediction rows)
- importer re-import reset deletes the image's prediction rows
The tagger_predictions JSON column is still dual-written (step 1) so it stays
valid during transition; the backfill task's NULL check still works. Removing
the JSON write + DROP column + retiring the #764 prune is the cleanup
follow-up (needs a quiesced-worker window for the DROP lock).
Tests: shared tests/_prediction_helpers.seed_predictions seeds the table;
read-path tests (suggestions, bulk consensus, allowlist apply, API) seed there
instead of ImageRecord.tagger_predictions.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Consumer #4 of the store-floor change (#764). An allowlist tag can't
auto-apply more permissively than the ingest floor — predictions below
tagger_store_floor aren't stored, so a lower min_confidence behaves
identically to the floor. update_threshold now clamps to max(value, floor);
the AllowlistTable confidence input min-binds to the live floor and clamps
on edit. Keeps the stored threshold honest about actual apply behavior.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Promotes the prediction store-floor from the TAGGER_STORE_FLOOR env (default
0.05) to a DB-backed, Settings-UI-tunable ml_settings column (default 0.70).
Storing every tag down to 0.05 from a ~10k-tag tagger is what grew
image_record's TOAST to ~100 GB; the suggestion path already filters at 0.70
and the centroid/learned path covers lower-confidence preferred tags, so the
sub-0.70 tail is redundant. Foundation for plan-task #764 (backfill + reclaim
land next; this only changes the write gate for NEW imports).
- ml_settings.tagger_store_floor (migration 0044, default 0.70)
- tagger.Tagger.infer(store_floor=...); ml task passes settings.tagger_store_floor
- ML admin GET/PATCH expose it; PATCH rejects a category suggestion threshold
below the floor (nothing below the floor is stored, so the gap surfaces
nothing) — server backstop for the UI slider clamp
- Settings → ML: store-floor slider + caption; category sliders min-bound to it
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The typed dropdown sourced the threshold-filtered panel list (>= 0.70 general),
so low-confidence actions/features the model DID predict never appeared — forcing
hand-typed custom tags instead of accepting the model's canonical formatting.
Add a threshold override: SuggestionService.for_image(threshold_override=) and
GET /images/<id>/suggestions?min=<f> surface EVERY stored prediction (down to the
0.05 store floor), alias-resolved and normalized, still excluding applied/rejected
and unsurfaced categories. The suggestions store gains allByCategory + loadAll
(min=0); the dropdown searches that full set (cap 20), while the Suggestions panel
stays curated at the configured threshold. Accept/dismiss drop from both lists.
Operator-asked 2026-06-09. Test: a 0.30 general prediction is hidden by default
but surfaced with threshold_override=0.0; unsurfaced categories still excluded.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Rule 8 'no letters -> drop' was over-eager: bare digit tags like '2005'
returned None even though they're legitimate (booru year-tag shape).
Widen the keep-condition to any alphanumeric. Emoticons (':/', '^_^',
'+_+') still drop since they contain neither letters nor digits.
Camie's booru-style vocab strings (`uchiha_sasuke_(naruto)`,
`#unicus_(idolmaster)`, `1000-nen_ikiteru_(vocaloid)`, `:/`) were
surfacing raw in SuggestionsPanel — and worse, the SAME raw string was
written to tag.name on Accept, polluting the DB with `underscored_lowercase`
names that don't match the operator's "Title Case" tag convention.
Add backend/app/services/ml/tag_name.py with a single normalize()
applying nine rules (strip leading junk #/./+/;/~/_/ws, drop trailing
_(disambiguator) blocks iteratively, strip wrapping quotes, underscores
to spaces, space after colon, title-case each word's first char,
preserve hyphens/apostrophes/digits, drop entries with no letters).
Wire into SuggestionService.for_image:
- raw Camie key kept for alias_map lookup (alias rows are hand-curated
against raw keys; don't disturb)
- display_name = normalize(raw); None means drop the candidate
- existing-tag lookup widened to case-insensitive match against BOTH
raw and normalized forms so legacy underscore-named Tag rows accepted
before this change still surface as "existing" not "+ new"
Closes the last two findings from the 2026-06-02 audit (G5.1 + G5.4).
G5.1 — Centroid version no longer drifts:
CentroidService now reads MLSettings.embedder_model_version (the DB
row tag_and_embed already writes from) for both the centroid model-
version stamp and the drift-detection comparison. Previously the
centroid sites imported MODEL_VERSION from env, so the version stamped
on centroids could disagree with the version stamped on the embeddings
they were built from. By construction those now match, so list_drifted
won't silently miss the env-vs-DB drift case.
embedder.py keeps MODEL_VERSION as an env-driven constant for the
actual model loader — that's a different concern (which weights are
loaded) from the version-stamp that gets persisted alongside data.
G5.4 — Modal is a Pinia-only overlay:
The previous URL↔modal sync in GalleryView and ArtistGalleryTab
leaked the modal across route changes (RouterLink to /artist/<slug>
left the modal mounted on top of the new route) and re-opened it
on history back/forward with stale ?image=N entries.
Now: openImage() just calls modal.open(id) — no URL push.
GalleryView's dead closeImage helper is deleted. A route.name
watcher in App.vue closes the modal whenever the route changes,
which auto-fixes RouterLink-in-modal and back/forward.
Backward-compat: ?image=N is still honored on initial mount as a
one-shot deep-link opener, then router.replace strips the query so
the URL doesn't re-trigger and no extra history entry is added.
Existing bookmarks / shared URLs keep working; new opens stay
Pinia-only.
Four coupled operator-asked changes to the view modal (Scribe plan #509):
1. **Autofocus tag entry on modal open** — TagAutocomplete grabs focus
in onMounted/nextTick so the caret is in the input the moment the
modal renders. No click needed to start typing.
2. **General suggestions expanded by default** — SuggestionsPanel's
general-category group now mounts with `:default-open="true"`.
Operator can collapse if too noisy, but the v1 frame shows them.
3. **Lower general threshold default 0.95 → 0.50** — MLSettings.
suggestion_threshold_general default matches character. Alembic
0029 also bumps the existing singleton row's value if it's still
at the old 0.95. Operator can re-tune from Settings → ML.
4. **Retire `copyright` + `artist` as ML suggestion categories** —
neither feeds a Tag.kind (`artist` retired in FC-2d-vii-c, never
really existed as a copyright tag-kind). They were surfaced in the
suggestions pipeline + threshold settings UI but had no follow-
through. Drop from SURFACED_CATEGORIES, suggestions._threshold_for,
ml_admin GET/PATCH allowlist, MLSettings columns (alembic 0029
drops the two columns), frontend CATEGORY_ORDER + CATEGORY_LABELS,
SuggestionsPanel.peopleCats, AliasPickerDialog kind-check, and
MLThresholdSliders rows.
Out of scope (intentional): `tag_kind` Postgres enum still includes
`artist` for historic Tag row queryability (per the model comment);
no operator pain reported, no enum-shrink needed.
Tests:
- test_surfaced_categories asserts {character, general}, excludes
artist + copyright.
- test_threshold_for_artist_is_unsurfaced extended to cover copyright.
- test_get_and_patch_settings asserts new 0.50 default and the absent
artist + copyright keys in the GET payload.
The HF repo Camais03/camie-tagger-v2 has camie-tagger-v2.onnx (789 MB)
+ camie-tagger-v2-metadata.json (7.77 MB) at root, NOT model.onnx +
selected_tags.csv. Tags ship as nested JSON (dataset_info.tag_mapping)
not CSV. Per the published onnx_inference.py reference: input is NCHW
not NHWC, normalize with ImageNet mean/std, pad-square color (124,116,
104), sigmoid the second output (refined predictions) not the first.
Operator hit this during the IR migration ML backfill — download_models
silently fetched only 3 json files (allow_patterns matched nothing
useful), tagger.load() then raised RuntimeError. Fetched the actual
v2 layout via WebFetch, rewrote tagger to match published reference.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
onnxruntime is in requirements-ml.txt only (deliberately kept out of the
lean web image and CI). The top-level `import onnxruntime` broke pytest
collection of test_ml_tagger / test_ml_suggestions / test_tasks_ml even
though those are pure-logic/integration-marked, because collection
imports the module.
Mirrors the embedder's lazy-torch pattern: onnxruntime is imported inside
Tagger.load(), placed AFTER the file-existence checks so
test_load_raises_when_model_missing still gets RuntimeError (not
ModuleNotFoundError) in onnxruntime-less environments. self._session
annotation dropped to a comment to avoid an eval-time ort reference.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
AllowlistService: accept (apply ml_accepted + add to allowlist + clear
rejection; returns whether newly-added so API can kick retro-apply),
add_alias_and_accept, dismiss, reject_applied_tag (remove + record
rejection so the allowlist won't re-apply), threshold update, remove,
list_all.
TagService.rename: refuses on (name, kind, fandom_id) collision with a
message pointing at FC-2c merge. Tests marked integration.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The read path: load tagger_predictions, drop unsurfaced categories
(rating/meta/year), apply per-category thresholds, batch-resolve aliases,
skip applied + rejected, augment with centroid hits above the similarity
threshold, merge duplicate signals (take max score, mark source 'both'),
group by category, sort by score DESC. Tests marked integration.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
recompute_for_tag (mean of member embeddings, eligible-kind + min-refs
gated, upsert), list_drifted (the delta-gate: member-count mismatch OR
missing OR wrong model version), find_similar_tags (pgvector cosine
distance, similarity = 1 - distance). Tests marked integration.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
resolve() / resolve_many() (batch, used by the suggestion read path),
idempotent create, remove, list_all. Category-scoped so 'naruto' as
character vs copyright map to different canonicals. Tests marked
integration (real DB).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Direct port of ImageRepo's siglip.py. Lazy torch/transformers import so
the web container can import the module (for enqueue logic) without the
torch cost. EMBED_DIM=1152 asserted against the schema's Vector(1152)
columns. Real inference runs in the local integration suite.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
CPU-only, lazy-loaded, process-singleton ONNX session. Parses Camie's
string-category selected_tags.csv (vs WD14's integer Danbooru ids).
STORE_FLOOR (0.05) keeps the stored predictions JSON compact;
SURFACED_CATEGORIES gates which categories the suggestion UI shows
(meta/rating/year stored but never surfaced).
Inference itself isn't unit-tested (1GB model not in CI); the missing-
model error path and pure-logic surface are. Full inference runs in the
local integration suite.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
bvandeusen