Files
silo-server/internal/recommendations/similar.go
T
eb6024573e feat(audiobooks): make audiobook libraries first-class catalog items (#73)
* docs(audiobooks): design spec for plugin absorption

Plan to absorb silo-plugin-audiobooks into silo-server as a first-party
feature. Audiobooks land in silo's existing SPA; ABS clients connect
directly. Hard constraints: reuse existing tables (media_items,
media_files, user_watch_progress, user_playback_sessions, people,
item_people, library_collections); only two new tables (abs_sessions,
podcast_feeds) and at most one column add (media_libraries.kind);
silo's main :8080 listener handles ABS Socket.io natively. Out of
scope: audiobook requests flow, smart collections, share links,
external recommender, custom metadata providers, separate audiobook
SPA.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* docs(audiobooks): implementation plan sub-plan 1 (discovery + schema)

First of six sub-plans for the absorption. Six tasks: a discovery
audit that resolves the spec's Risk questions, four idempotent SQL
migrations (abs_sessions, podcast_feeds, media_libraries.kind,
audiobooks.enabled feature flag), and an empty-but-compiling
internal/audiobooks package scaffolded into cmd/silo. Lands as a
strict no-op for users (feature flag defaults to false).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* docs(audiobooks): discovery findings for absorption sub-plan 1

Locks schema/code decisions for migrations 139-142 and downstream
sub-plans. Resolves open Risk questions from the absorption design spec.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): migration 139 add abs_sessions table

Parallel of jellycompat_sessions for Audiobookshelf-compatible clients.
Lets ABS mobile/desktop apps maintain a device-bound session that
silo's audiobooks/abs handlers will validate.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* style(audiobooks): match codebase conventions in migration 139

Lowercases type keywords in the abs_sessions CREATE TABLE body to
match neighboring migrations, fixes the client_version column
alignment, and replaces the misleading "parallel to
jellycompat_sessions" header comment with a more accurate
description of the table's role.

Cosmetic only — the running schema is unchanged.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): migration 140 add podcast_feeds table

Side table on media_items for RSS-subscribed podcasts. Holds feed URL,
ETag/Last-Modified for conditional fetches, last-refresh timestamp, and
the per-feed refresh interval consumed by the upcoming
podcastfeed.Refresher scheduled task.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* style(audiobooks): uppercase PRIMARY KEY in migration 140

Aligns with the codebase convention (type keywords lowercase,
constraint keywords uppercase) established in migration 139's
post-style-fix form. Cosmetic only — running schema is unchanged.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* chore(audiobooks): migration 141 no-op for media_folders.type

Sub-plan 1 originally reserved migration 141 to add a 'kind' column to
media_libraries discriminating audiobook/podcast libraries. Discovery
audit (sub-plan 1 Task 1) found that the actual table is media_folders
and it already has a type text NOT NULL column with no CHECK constraint
or enum, so 'audiobooks' and 'podcasts' can be added as future values
without DDL.

Landing this migration as a documented no-op preserves the version
numbering audit trail and pins the decision in git history. The
matching down migration is also a no-op.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): migration 142 add audiobooks.enabled flag

Server-settings row that gates the absorbed audiobooks feature.
Defaults to 'false' so sub-plan 1 lands as a strict no-op; subsequent
sub-plans branch on this flag and operators flip it to 'true' at
cutover.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): scaffold internal/audiobooks package

Empty-but-compiling Service that reads the audiobooks.enabled feature
flag from server_settings. Wired into cmd/silo so the package is
referenced from the binary; no routes mounted, no scheduled tasks
registered, no DB writes. Subsequent sub-plans hang scanner branches,
ABS handlers, Socket.io, podcast refresher, and SPA pages off this
Service.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* style(audiobooks): cosmetic cleanups in scaffolded package

Two pre-emptive cleanups flagged by code review before sub-plan 2
copies the patterns:

  1. Sort the internal/audiobooks import after internal/adminjob in
     cmd/silo/main.go (alphabetical).
  2. Drop the redundant "audiobooks: " prefix from the Enabled() error
     wrap; matches how every other top-level service package
     (watchstate, scanqueue, metadata, etc.) formats errors.

No behavior change.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* docs(audiobooks): implementation plan sub-plan 2 (scanner)

Second of six sub-plans. 10 tasks: PersonKind constants for Author and
Narrator, audio-extension recognizer, library-type helpers, a
walkLogicalTree refactor (movieLibrary bool -> typed walkMode), chapter
extraction via ffprobe, single-file and multi-file audiobook parsers,
scanner write path producing media_items.type='audiobook', and a
filesystem podcast parser (RSS deferred to sub-plan 5).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): add Author and Narrator PersonKind constants

Discovery audit confirmed item_people.kind is unconstrained smallint
with values 1-6 in use. Reserve 7 = Author, 8 = Narrator for audiobook
people-links written by the upcoming scanner branches.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): add audio-extension recognizer for scanner

Mirrors the existing videoExtensions/SupportsVideoFile pair. Used by
upcoming audiobook and podcast scanner branches to filter directory
walks.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): library-type recognizers for scanner dispatch

isAudiobookLibraryType and isPodcastLibraryType match singular and
plural forms case-insensitively, mirroring isMovieLibraryType. Used by
upcoming scanner walk branches (Task 4) that filter audio files into
audiobook and podcast libraries.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* refactor(scanner): replace movieLibrary bool with typed walkMode

Lets walkLogicalTree dispatch on multiple library shapes (video, movie,
audiobook, podcast) without proliferating boolean flags. Behavior for
existing video and movie libraries is unchanged; audiobook and podcast
modes will be consumed by the upcoming audiobook.go and podcast.go
parsers in later tasks of this sub-plan.

walkModeFor() derives the mode from a media_folders.type string;
unknown types default to walkModeVideo to preserve prior behavior for
any caller still passing a raw type.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): expose ffprobe format tags on ProbeData

The audiobook scanner needs format-level tags (title, artist, album,
date) for media_items metadata; ffprobe already parses them in
ffprobeFormat.Tags but ProbeData previously discarded them. Add
FormatTags map[string]string to ProbeData, populate it in
convertProbeData via a new normalizeFormatTags helper that lowercases
keys and trims values.

Adds a fixture audiobook .m4b with embedded chapters (Intro/Outro) and
format tags, and a test that verifies ProbeFile() returns both
correctly.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): parser for single-file audiobook folders

parseAudiobookFolder reads tags + chapters via the existing ProbeFile
(now that Task 5 exposes FormatTags on ProbeData) and produces a
parsedAudiobook struct. Title falls back from "title" tag to "album";
author from "artist" -> "album_artist" -> "composer"; series from
"album" -> "series" -> "mvnm" (Movement Name, used by some MP4 tools).
Year parsed from "date" or "year" tags, tolerating ISO dates and
parenthesized forms.

Single-file case only; multi-file folders (one audio file per chapter)
return a placeholder error and arrive in Task 7.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): multi-file audiobook folder support

Folders containing N audio files (one per chapter/part) get one
parsedAudiobookFile per file; each file's chapter list is synthesized
as a single chapter with title = filename stem. Title/author/series/
year come from the first file's tags.

Also drops the duplicate pickFirstNonEmpty helper added in Task 6 in
favor of the existing firstNonEmpty already in probe.go.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): scanner write path produces audiobook media_items

ScanAudiobookFolder walks an audiobooks-typed media folder and treats
each immediate subdirectory as one audiobook. For each parsed audiobook
it upserts:
  - one media_items row with type='audiobook'
  - one media_files row per audio file (with chapters JSONB)
  - author/narrator links in item_people (kind=7, kind=8)

Adds itemRepo and personRepo to the Scanner struct, wired from
fileRepo.Pool() in NewScanner — no constructor signature change needed.

ScanFolder dispatches to this path when folder.Type='audiobooks',
bypassing the per-file movie/TV pipeline because audiobooks are
folder-scoped entities.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): filesystem podcast scanner

ScanPodcastFolder walks a podcasts-typed media folder, treating each
subdirectory as a podcast show and each audio file inside as an
episode. Writes media_items.type='podcast' + episodes rows + media_files
rows. RSS-subscribed feeds (podcast_feeds table) arrive in sub-plan 5;
this task covers filesystem-only ingestion.

ScanFolder dispatches to this path when folder.Type='podcasts'.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* docs(audiobooks): implementation plan sub-plan 5 (podcasts)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): expose audiobooks/podcasts library types in admin UI

Adds 'Audiobooks' and 'Podcasts' options to the library-type dropdown
in the admin libraries page so operators can flag a folder as an
audiobook or podcast library. Extends contentLevelsForType() so the
admin UI's downstream filtering treats those types correctly
(audiobook -> ['audiobook'], podcasts -> ['podcast',
'podcast_episode']).

Backend scanner branches for these types were already wired in
sub-plan 2.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* chore(migrations): renumber 139_abs_sessions to 147 for origin/main merge

origin/main adds 139_media_requests at the same number our local
audiobook branch had used for abs_sessions. Renumber ours to 147 to
free up 139 for the upstream migration. The schema_versions row is
updated in lockstep on the running database so the migrator sees the
abs_sessions migration as already applied at its new version.

Migrations 140-146 (podcast feeds, media_folders kind noop, audiobook
feature flag, abs playback sessions, podcast episode guid, audiobook
series, audiobook title cleanup) stay where they are — they don't
collide with anything on origin/main.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* chore(migrations): renumber 140_podcast_feeds to 157 for origin/main merge

origin/main added 140_user_permissions at the same version this branch
had used for podcast_feeds. Renumber ours to 157 (next free above the
collections-unify migration at 156) so 140 is free for the upstream
migration. schema_versions on the running database is updated in lockstep
so the migrator sees podcast_feeds as already applied at its new version.

Same pattern as d59c1cb (renumber 139_abs_sessions to 147 for the prior
main merge). Pending migrations after this rename: 132 (downloaded
subtitles admin index, main), 140 (user_permissions, main), and 156
(unify_user_collections, this branch).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* chore(migrations): renumber 141_media_folders_kind_noop to 159 for origin/main merge

Same shape as eb8f67d (the 140→157 renumber from the previous main
merge). origin/main added 141_episode_title_sort_index at the same
version this branch had used for media_folders_kind_noop. Renumber
ours to 159 (next free above the audiobook_series truncate at 158) so
141 is open for the upstream migration. schema_versions on the
running database is updated in lockstep so the migrator sees
media_folders_kind_noop as already applied at its new version.

Pending migrations on silo-prod after this rename: 141
(episode_title_sort_index, main) and any other newer ones from main
that the branch hasn't picked up yet.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* chore(migrations): renumber 142_audiobooks_feature_flag to 160 for origin/main merge

Companion to 3c6f062's 141 renumber — origin/main also added
142_episode_catalog_entries (alongside 141_episode_title_sort_index)
at a version this branch had used for the audiobooks feature flag.
Renumber ours to 160 so 142 is open for the upstream migration;
schema_versions on silo-prod is updated in lockstep so the migrator
sees audiobooks_feature_flag as already applied at its new version.

This was the only remaining collision (verified by checking for
duplicate version prefixes across migrations/).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(audiobooks): address foundation review comments

* fix(audiobooks): tighten scanner identity handling

* fix(audiobooks): propagate scanner cancellation

* chore(audiobooks): adopt goose migration layout

* docs(audiobooks): implementation plan sub-plan 3 (API + frontend MVP)

Third of six sub-plans. 9 tasks: three REST endpoints (list/detail/
progress), TanStack Query hooks + types, three React pages
(Library/Detail/Player), and navigation integration. Scoped to MVP —
author/series indices, smart collections, share links, and other
nice-to-haves from the spec are deferred. Streaming reuses silo's
existing /api/v1/stream/{session_id}; no new transcode code.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): list endpoint at GET /api/v1/audiobooks

Paginated list of media_items with type='audiobook' scoped to the
caller's accessible libraries via the existing access filter.
Mirrors silo's existing list-style handlers for movies and series.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): detail endpoint at GET /api/v1/audiobooks/{id}

Returns the media_items row, its media_files (with chapters JSONB),
author/narrator extracted from item_people (kinds 7/8), and the
caller's per-profile listening progress from user_watch_progress.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): progress endpoint at POST /api/v1/audiobooks/{id}/progress

UPSERTs user_watch_progress for the caller's (user_id, profile_id,
content_id). Body carries position_seconds; clients are expected to
post every 5-10s during playback plus on pause/seek (matching silo's
existing video progress cadence).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): frontend types and TanStack Query hooks

TypeScript types match the JSON shapes from the new
/api/v1/audiobooks endpoints (list, detail, progress). Three hooks:
useAudiobookLibrary (list), useAudiobook (detail), and
useReportAudiobookProgress (mutation that invalidates the detail
query on success so progress updates reflect immediately).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): library grid page at /audiobooks

Renders a paginated grid of audiobook cards using the
useAudiobookLibrary hook. Each card links to /audiobooks/book/{id}.
Cards show poster, title, and year; falls back to a "No cover"
placeholder when the audiobook has no poster_url. Empty state hints
to operators that they need to set a library's type to 'audiobooks'.

Routes themselves are wired in Task 8 (navigation integration).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): detail page with chapter list

Renders cover, title, author, narrator, year, and overview alongside a
chapter list. Clicking a chapter opens an inline sticky
AudiobookPlayer at that chapter's start. A "Resume" button restarts
playback at the saved progress position if present.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): HTML5 audio player with chapter navigation

Single-file audiobook playback for MVP. Multi-file queuing arrives in
a follow-up. Streams via the existing /api/v1/direct-download GET
endpoint. Position is reported to /api/v1/audiobooks/{id}/progress
every 10s while playing plus on pause/seek/end. Skip-30s, playback
rate select, chapter list panel.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(audiobooks): wire navigation and routes

Adds an Audiobooks entry to the sidebar and registers the two new
routes (/audiobooks for the library grid, /audiobooks/book/:id for
detail). The player renders inline inside the detail page; no
dedicated player route is required for MVP.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(audiobooks): address native API review comments

* feat(audiobooks): add ABS compatibility and polish

* fix(audiobooks): stabilize ABS playback progress reporting

* fix(audiobooks): clean up ABS branch review fixes

* chore(audiobooks): adopt goose layout for ABS migrations

* fix(audiobooks): align player seek bar props

* feat(audiobooks): make libraries first-class catalog items

* feat(admin): add server restart endpoint

* fix(audiobooks): address review comment findings

---------

Co-authored-by: RXWatcher <14085001+RXWatcher@users.noreply.github.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-06-07 15:57:05 -04:00

420 lines
12 KiB
Go

package recommendations
import (
"context"
"fmt"
"log/slog"
"strings"
"github.com/Silo-Server/silo-server/internal/models"
"github.com/Silo-Server/silo-server/internal/recommendations/embeddings"
)
// isQuotaError returns true if the error indicates an API quota/billing issue
// that won't be resolved by retrying.
func isQuotaError(err error) bool {
msg := err.Error()
return strings.Contains(msg, "insufficient_quota") ||
strings.Contains(msg, "exceeded your current quota") ||
strings.Contains(msg, "billing")
}
func embeddingTextNeedsRefresh(storedModel, storedCanonicalText, generatedCanonicalText, currentModel string) bool {
return storedModel != currentModel || storedCanonicalText != generatedCanonicalText
}
// SimilarItems returns items most similar to the given item. Blends embedding
// similarity (70%) with co-watch Jaccard score (30%), applies a validation
// pipeline, MMR re-ranking, and assigns connection reasons.
func (e *Engine) SimilarItems(ctx context.Context, itemID string, limit int) ([]ScoredItem, error) {
// 1. Fetch source embedding.
embedding, err := e.repo.GetEmbedding(ctx, itemID)
if err != nil {
return nil, fmt.Errorf("get embedding for item %s: %w", itemID, err)
}
if embedding == nil {
return nil, nil
}
// 2. Fetch source metadata for validation pipeline.
sourceMeta, err := e.repo.GetItemMetadata(ctx, itemID)
if err != nil {
slog.Warn("could not fetch source metadata, skipping validation", "item_id", itemID, "error", err)
}
sourceType := ""
if sourceMeta != nil {
sourceType = sourceMeta.Type
}
// 3. Embedding search (3x limit for filtering headroom). Constrain to the
// source item's media type so an audiobook never appears in a movie's
// Similar rail (and vice versa) once audiobook embeddings exist.
embCandidates, err := e.repo.FindSimilar(ctx, embedding, []string{itemID}, sourceType, limit*3)
if err != nil {
return nil, fmt.Errorf("find similar items: %w", err)
}
// 4. Co-watch neighbors.
cowatchPairs, _ := e.repo.GetCowatchNeighbors(ctx, itemID, limit*3)
cowatchMap := make(map[string]float64, len(cowatchPairs))
for _, p := range cowatchPairs {
cowatchMap[p.SimilarItemID] = p.JaccardScore
}
// 5. Blend scores (70% embedding, 30% co-watch).
blended := blendScores(embCandidates, cowatchMap, 0.7, 0.3)
// 6. Validation pipeline.
if sourceMeta != nil && len(blended) > 0 {
blended = e.applyValidation(ctx, sourceMeta, blended)
}
// 7. MMR re-ranking.
candidateIDs := make([]string, len(blended))
for i, item := range blended {
candidateIDs[i] = item.MediaItemID
}
embMap, _ := e.repo.GetBatchEmbeddings(ctx, candidateIDs)
result := applyMMR(blended, embMap, e.mmrLambda(LambdaSimilarItems), limit)
// 8. Connection reasons.
e.assignReasons(ctx, itemID, sourceMeta, result)
return result, nil
}
// applyValidation filters and penalizes candidates using the validation pipeline.
func (e *Engine) applyValidation(ctx context.Context, sourceMeta *ItemMetadata, candidates []ScoredItem) []ScoredItem {
candidateIDs := make([]string, len(candidates))
for i, item := range candidates {
candidateIDs[i] = item.MediaItemID
}
// Batch fetch candidate genres and titles.
candidateGenres, _ := e.repo.GetItemAllGenres(ctx, candidateIDs)
// We need titles for title-pattern detection. Fetch lightweight metadata.
type titleYear struct {
title string
year int
}
candTitles := make(map[string]titleYear)
rows, err := e.pool.Query(ctx, `
SELECT content_id, title, COALESCE(year, 0) FROM media_items
WHERE content_id = ANY($1)`, candidateIDs)
if err == nil {
defer rows.Close()
for rows.Next() {
var id, title string
var year int
if rows.Scan(&id, &title, &year) == nil {
candTitles[id] = titleYear{title: title, year: year}
}
}
}
var validated []ScoredItem
for _, item := range candidates {
candMeta := &ItemMetadata{
Genres: candidateGenres[item.MediaItemID],
}
if ty, ok := candTitles[item.MediaItemID]; ok {
candMeta.Title = ty.title
candMeta.Year = ty.year
}
result := validateCandidate(sourceMeta, candMeta, item)
if result.rejected {
continue
}
if result.scoreMult < 1.0 {
item.Score *= result.scoreMult
}
validated = append(validated, item)
}
return validated
}
// assignReasons computes connection reasons for each result item.
func (e *Engine) assignReasons(ctx context.Context, sourceItemID string, sourceMeta *ItemMetadata, items []ScoredItem) {
if len(items) == 0 {
return
}
// Default all to similar_content.
for i := range items {
items[i].Reason = "similar_content"
}
if sourceMeta == nil {
return
}
candIDs := make([]string, len(items))
for i := range items {
candIDs[i] = items[i].MediaItemID
}
// Batch fetch people for source + candidates.
allIDs := make([]string, 0, len(candIDs)+1)
allIDs = append(allIDs, sourceItemID)
allIDs = append(allIDs, candIDs...)
allPeople := make(map[string]*itemPeopleData)
if e.personRepo != nil {
peopleMap, err := e.personRepo.ListForItems(ctx, allIDs)
if err == nil {
for id, people := range peopleMap {
pd := &itemPeopleData{}
for _, p := range people {
switch p.Kind {
case models.PersonKindDirector:
pd.directors = append(pd.directors, p.Name)
case models.PersonKindActor:
if p.SortOrder <= 5 {
pd.actors = append(pd.actors, p.Name)
}
}
}
allPeople[id] = pd
}
}
}
// Batch fetch candidate studios and genres for reasons.
candStudios := make(map[string][]string)
candidateGenres, _ := e.repo.GetItemAllGenres(ctx, candIDs)
rows, err := e.pool.Query(ctx, `
SELECT content_id, studios FROM media_items
WHERE content_id = ANY($1) AND studios IS NOT NULL`, candIDs)
if err == nil {
defer rows.Close()
for rows.Next() {
var id string
var studios []string
if rows.Scan(&id, &studios) == nil {
candStudios[id] = studios
}
}
}
sourcePeople := allPeople[sourceItemID]
if sourcePeople == nil {
sourcePeople = &itemPeopleData{}
}
for i := range items {
candPeople := allPeople[items[i].MediaItemID]
if candPeople == nil {
candPeople = &itemPeopleData{}
}
candGenres := candidateGenres[items[i].MediaItemID]
reason, detail := computeReason(sourcePeople, candPeople, sourceMeta.Studios, candStudios[items[i].MediaItemID], sourceMeta.Genres, candGenres)
items[i].Reason = reason
items[i].ReasonDetail = detail
}
}
// EmbedItem generates and stores an embedding for a single media item.
func (e *Engine) EmbedItem(ctx context.Context, itemID string) error {
if err := e.ensureEmbeddingLockConfig(ctx); err != nil {
return fmt.Errorf("embed item %s: %w", itemID, err)
}
items, err := e.itemRepo.GetByIDs(ctx, []string{itemID})
if err != nil || len(items) == 0 {
return fmt.Errorf("get item %s: %w", itemID, err)
}
// Hydrate cast/crew from item_people for richer embedding text.
e.hydrateItemPeople(ctx, items)
text := embeddings.BuildEmbeddingText(items[0])
vectors, err := e.embClient.Embed(ctx, []string{text})
if err != nil {
return fmt.Errorf("embed item %s: %w", itemID, err)
}
if len(vectors) == 0 {
return fmt.Errorf("no embedding returned for item %s", itemID)
}
if err := e.ensureEmbeddingLock(ctx, vectors[0]); err != nil {
return fmt.Errorf("embed item %s: %w", itemID, err)
}
return e.repo.UpsertEmbedding(ctx, itemID, vectors[0], e.cfg.EmbeddingModel, text)
}
// EmbedAll embeds all items that are missing embeddings or have stale canonical text.
func (e *Engine) EmbedAll(ctx context.Context) (int, error) {
model := e.cfg.EmbeddingModel
total := 0
batchSize := 10
afterID := ""
if err := e.ensureEmbeddingLockConfig(ctx); err != nil {
return total, err
}
for {
select {
case <-ctx.Done():
return total, ctx.Err()
default:
}
candidates, err := e.repo.ListEmbeddingTextCandidates(ctx, afterID, model, batchSize)
if err != nil {
return total, err
}
if len(candidates) == 0 {
break
}
afterID = candidates[len(candidates)-1].MediaItemID
ids := make([]string, 0, len(candidates))
stored := make(map[string]EmbeddingTextCandidate, len(candidates))
for _, candidate := range candidates {
ids = append(ids, candidate.MediaItemID)
stored[candidate.MediaItemID] = candidate
}
items, err := e.itemRepo.GetByIDs(ctx, ids)
if err != nil {
return total, fmt.Errorf("get items for embedding: %w", err)
}
// Hydrate cast/crew for richer embedding text.
e.hydrateItemPeople(ctx, items)
texts := make([]string, len(items))
staleItems := make([]*models.MediaItem, 0, len(items))
staleTexts := make([]string, 0, len(items))
for i, item := range items {
texts[i] = embeddings.BuildEmbeddingText(item)
candidate := stored[item.ContentID]
if embeddingTextNeedsRefresh(candidate.Model, candidate.CanonicalText, texts[i], model) {
staleItems = append(staleItems, item)
staleTexts = append(staleTexts, texts[i])
}
}
if len(staleItems) == 0 {
continue
}
vectors, err := e.embClient.Embed(ctx, staleTexts)
if err != nil {
// Quota/billing errors won't resolve by retrying — stop immediately.
if isQuotaError(err) {
return total, fmt.Errorf("embedding API quota exceeded, check billing: %w", err)
}
// Batch failed — fall back to embedding one at a time so a single
// oversized item doesn't block the entire job.
slog.Warn("batch embed failed, falling back to single-item mode", "error", err, "batch_size", len(staleItems))
for i, item := range staleItems {
if ctx.Err() != nil {
return total, ctx.Err()
}
single := staleTexts[i]
vecs, embedErr := e.embClient.Embed(ctx, []string{single})
if embedErr != nil {
if isQuotaError(embedErr) {
return total, fmt.Errorf("embedding API quota exceeded, check billing: %w", embedErr)
}
if ctx.Err() != nil {
return total, ctx.Err()
}
slog.Warn("skipping item, embed failed", "item_id", item.ContentID, "error", embedErr)
continue
}
if len(vecs) > 0 {
if err := e.ensureEmbeddingLock(ctx, vecs[0]); err != nil {
return total, fmt.Errorf("embed item %s: %w", item.ContentID, err)
}
if storeErr := e.repo.UpsertEmbedding(ctx, item.ContentID, vecs[0], model, single); storeErr != nil {
slog.Warn("skipping item, store failed", "item_id", item.ContentID, "error", storeErr)
continue
}
total++
}
}
continue
}
for i, item := range staleItems {
if i >= len(vectors) {
break
}
if err := e.ensureEmbeddingLock(ctx, vectors[i]); err != nil {
return total, fmt.Errorf("embed item %s: %w", item.ContentID, err)
}
if err := e.repo.UpsertEmbedding(ctx, item.ContentID, vectors[i], model, staleTexts[i]); err != nil {
slog.Warn("skipping item, store failed", "item_id", item.ContentID, "error", err)
continue
}
total++
}
}
return total, nil
}
// hydrateItemPeople populates the People field on media items from item_people.
func (e *Engine) hydrateItemPeople(ctx context.Context, items []*models.MediaItem) {
ids := make([]string, len(items))
for i, item := range items {
ids[i] = item.ContentID
}
peopleMap, err := e.personRepo.ListForItems(ctx, ids)
if err != nil {
slog.Warn("failed to hydrate item people for embeddings", "error", err)
return
}
for _, item := range items {
if people, ok := peopleMap[item.ContentID]; ok {
item.People = people
}
}
}
func (e *Engine) ensureEmbeddingLock(ctx context.Context, vector []float32) error {
sourceDimensions := len(vector)
lock, err := e.repo.GetEmbeddingLock(ctx)
if err != nil {
return fmt.Errorf("load embedding lock: %w", err)
}
if lock == nil {
return e.repo.SetEmbeddingLock(ctx, EmbeddingLock{
BaseURL: e.cfg.EmbeddingBaseURL,
Model: e.cfg.EmbeddingModel,
SourceDimensions: sourceDimensions,
StorageDimensions: CanonicalEmbeddingDimensions,
})
}
if err := lock.Validate(e.cfg.EmbeddingBaseURL, e.cfg.EmbeddingModel, sourceDimensions); err != nil {
return err
}
return nil
}
func (e *Engine) ensureEmbeddingLockConfig(ctx context.Context) error {
lock, err := e.repo.GetEmbeddingLock(ctx)
if err != nil {
return fmt.Errorf("load embedding lock: %w", err)
}
if lock == nil {
return nil
}
if err := lock.ValidateConfig(e.cfg.EmbeddingBaseURL, e.cfg.EmbeddingModel); err != nil {
return err
}
return nil
}