mirror of
https://github.com/got-feedBack/feedBack.git
synced 2026-07-22 21:01:40 +00:00
fix/plugin-esc-global
4 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
ChrisBeWithYou
|
74cff4e0d6
|
feat(enrichment): alias-aware scoring — auto-confirm non-Latin-primary artists (#772)
Some checks are pending
ship-ci / ci (push) Waiting to run
* feat(enrichment): loose MusicBrainz search fallback (find aliased artists)
The MB text search used a strict field-phrase query
(`recording:"<title>" AND artist:"<artist>"`). A field phrase only matches
MusicBrainz's *primary* artist/title — it never searches ALIASES — so a
recording stored under a non-Latin primary name (大橋純子) whose romanized
form ("Junko Ohashi") is only an alias returns ZERO results, even though MB
has it. Whole swaths of a community library (e.g. romanized J-pop / city-pop
charts) were unsearchable.
- `build_recording_query(..., loose=True)` drops the field scoping + phrases
for plain AND-ed term groups (`(telephone number) AND (junko ohashi)`),
which searches the whole document incl. aliases.
- `_mb_search_recordings` runs the strict query first (unchanged, high
precision) and only on an EMPTY result retries once with the loose query —
so mainstream matches are untouched and the extra throttled request is spent
only on a miss. Results are re-scored by rank_candidates, so recall goes up
without lowering match quality (auto-accept still needs the per-field floors).
Verified live: "Junko Ohashi / Telephone Number" and "Anri / Windy Summer"
(both 0 under the strict query) now surface the real records; "AC/DC /
Highway to Hell" still hits strict at score 1.0 with no loose retry.
Follow-up (separate): alias-aware SCORING so these can auto-confirm, not just
appear as manual candidates.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(enrichment): alias-aware scoring (auto-confirm non-Latin-primary artists)
Builds on the loose-search fallback: that surfaces a recording stored under a
Japanese primary name (大橋純子) via its romanized alias, but the SCORER still
compared the reference ("Junko Ohashi") against the primary only → artist
similarity 0 → below the auto floor, so it could only ever be a manual
candidate, never an auto-fill.
- mb_match: `cand_artist_sim` takes the best similarity over the candidate's
primary name AND its `artist_aliases`; score_candidate + classify use it.
- server: `_mb_artist_aliases(id)` fetches an artist's aliases (one throttled
lookup, process-cached — a one-artist discography costs ONE request) and
`_alias_enrich` attaches them ONLY to promising near-misses (title agrees,
primary artist doesn't) so a normal pass spends zero extra requests. Wired
into both the auto-matcher (_enrich_one) and the manual search proxy.
Verified live: "Junko Ohashi / Telephone Number" → 大橋純子 candidate goes from
score 0.5 (loose-only) to 1.0 (auto-confirmable), ranked #1; "AC/DC / Highway
to Hell" unchanged at 1.0 with no alias lookup.
Stacks on #771 (feat/mb-loose-search-fallback).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(enrichment): keep live exclusion in the loose search fallback
The loose fallback dropped the strict path's -secondarytype:Live filter, so a
studio chart whose strict query missed could fall back to — and, since
score_candidate doesn't penalize live takes, auto-confirm — a live-only
recording. Apply the same live gate to the loose query (skipped only when the
source title is itself a live take).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: byrongamatos <xasiklas@gmail.com>
|
||
|
ChrisBeWithYou
|
18c4e229e1
|
feat(enrichment): loose MusicBrainz search fallback (find aliased/romanized artists) (#771)
* feat(enrichment): loose MusicBrainz search fallback (find aliased artists)
The MB text search used a strict field-phrase query
(`recording:"<title>" AND artist:"<artist>"`). A field phrase only matches
MusicBrainz's *primary* artist/title — it never searches ALIASES — so a
recording stored under a non-Latin primary name (大橋純子) whose romanized
form ("Junko Ohashi") is only an alias returns ZERO results, even though MB
has it. Whole swaths of a community library (e.g. romanized J-pop / city-pop
charts) were unsearchable.
- `build_recording_query(..., loose=True)` drops the field scoping + phrases
for plain AND-ed term groups (`(telephone number) AND (junko ohashi)`),
which searches the whole document incl. aliases.
- `_mb_search_recordings` runs the strict query first (unchanged, high
precision) and only on an EMPTY result retries once with the loose query —
so mainstream matches are untouched and the extra throttled request is spent
only on a miss. Results are re-scored by rank_candidates, so recall goes up
without lowering match quality (auto-accept still needs the per-field floors).
Verified live: "Junko Ohashi / Telephone Number" and "Anri / Windy Summer"
(both 0 under the strict query) now surface the real records; "AC/DC /
Highway to Hell" still hits strict at score 1.0 with no loose retry.
Follow-up (separate): alias-aware SCORING so these can auto-confirm, not just
appear as manual candidates.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(enrichment): keep live exclusion in the loose search fallback
The loose fallback dropped the strict path's -secondarytype:Live filter, so a
studio chart whose strict query missed could fall back to — and, since
score_candidate doesn't penalize live takes, auto-confirm — a live-only
recording. Apply the same live gate to the loose query (skipped only when the
source title is itself a live take).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: byrongamatos <xasiklas@gmail.com>
|
||
|
ChrisBeWithYou
|
a65d8cfa13
|
fix(enrichment): rank the canonical studio take over live/comp versions (#758)
* fix(enrichment): rank the canonical studio take over live/comp versions A flat MusicBrainz /recording text search ties every take of a song at the same score, so "AC/DC — Highway to Hell" returns a wall of live bootlegs and compilations with the 1979 studio version buried (or below the fetch limit). - build_recording_query: drop live-ONLY recordings (`-secondarytype:Live`). Compilations are deliberately kept — they REUSE the studio recording, so filtering them cuts the very recording we want (verified against MB). - _best_release / parse_recording_doc: pick the canonical studio album (primary Album, no Live/Compilation/Remix/... secondary type) for the displayed album/year, and expose a `studio` flag. - rank_candidates: since the combined score caps at 1.0 (perfect text match ties), break ties on the studio flag and — when the caller knows the audio length — on duration proximity, so the studio take wins over live/extended cuts. The studio distinction is intentionally NOT scored (a live take is still the right SONG), only re-ordered. - /api/enrichment/search: accept an optional `duration` param so a caller that has the audio but no library row (the editor's create modal) can pass the master-track length for the duration tiebreak. Verified end-to-end against live MusicBrainz: AC/DC "Highway to Hell" now returns the 1979 studio recording at #1 with the correct album + year. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(enrichment): official releases outrank unofficial studio albums _best_release sorted (clean, status_ok, date), so an UNofficial bootleg Album outranked an official Single/EP/comp — regressing canonical album/year and seeding cover-art from a bootleg for single-only songs. Order status_ok before clean: official first, then prefer a clean studio album among the official releases (still surfaces the studio album over an official live/comp album). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(enrichment): keep live recordings for genuinely-live charts build_recording_query unconditionally added -secondarytype:Live, but denoise() strips a '(Live at …)' qualifier from the query — so a chart that IS a live take had its only correct recording filtered out (both background enrichment and manual search). Skip the live filter when the source title carries a parenthetical live marker; a bare title word ('Live and Let Die') still filters. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(enrichment): drop the studio tiebreak when the chart is a live take Follow-through on keeping live recordings for live charts: rank_candidates still ranked the studio take ahead of a tied live one, so a live chart would auto-match the studio recording. Skip the studio tiebreak when the source title has a live marker — duration proximity + score then pick the right live version. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> |
||
|
ChrisBeWithYou
|
74cd08f765
|
library: MusicBrainz text matching + Match-Review UI (P8) (#710)
* library: MusicBrainz text matching + Match-Review UI — P8
Replaces the enrichment plumbing's no-op matcher (P7) with the real
pipeline, per the library-metadata design: a wrong match is worse than a
slow one, so medium confidence goes to a human review queue and never
straight to canonical values.
- lib/mb_match.py (new, pure — no network/DB/server imports): denoise
(author credits, (440Hz)/(Live)/(No Lead)/(v2) parentheticals,
diacritics/punctuation, ACDC / AC DC / AC/DC folding via compacted
token equality), token-set similarity, scoring with year/duration
corroboration bonuses, tier classification (auto needs combined
>= 0.95 AND per-field floors — a perfect-title cover by the wrong
artist, or a chart with no artist, can never auto-match), Lucene
query building, MusicBrainz response normalization.
- Matcher precedence in _enrich_one: content-hash cache copy (another
chart of the same recording matches with no network) -> manifest
mbid (tier 0) / isrc (tier 1) exact keys, feature-detected and
strictly shape-validated, read-only -> text search tiers
(auto / review / failed).
- Lifecycle: review rows store their ranked candidate list (JSON) and
write NO canonical fields until a human accepts; failed rows retry on
an exponential backoff (1 h doubling, 7 d cap) via the attempts
column; user-rejected rows never auto-retry; an identity edit
re-queues anything and resets the backoff; never-overwrite-manual is
enforced inside the single writer (apply_enrichment_match) so no call
path can forget it.
- Network: _mb_http_get is the one transport seam — throttled to
<= 1 req/s through P7's _enrich_throttle, identified with a real
User-Agent from VERSION, and a 503 pauses the whole pass without
burning attempts. Offline guard: no sockets under
FEEDBACK_ENRICH_OFFLINE or FEEDBACK_SKIP_STARTUP_TASKS, so pytest can
never reach MusicBrainz; the pass still stamps identity hashes
(two-phase), which is why every P7 test passes unchanged.
- Routes: GET /api/enrichment/review, POST
/api/enrichment/review/{filename}/accept|reject|pick, GET
/api/enrichment/search (throttled manual-search proxy). All four are
demo-mode blocked.
- Match facet: match= CSV accepted by /api/library AND
/api/library/stats (the A-Z rail's letter counts stay lockstep with
the grid) — review / matched (incl. manual) / unmatched / pending,
the same EXISTS idiom as the mastery facet.
- UI: static/v3/match-review.js (new, self-contained) — an ambient
"N to review" chip beside the song count (rendered only when
non-zero; silent on success, no toasts), and a review drawer on the
filter-drawer slide idiom (Escape + focus trap; row click accepts,
"Not a match" rejects, "Search instead" is the fix-match escape
hatch). songs.js gets the chip mount, a Match filter section, and
session-only match state; also fixes the latent applySavedPrefs bug
where restored filters dropped the mastery key, which made the
filter drawer throw for anyone with saved prefs.
- static/tailwind.min.css regenerated (scripts/build-tailwind.sh) for
the new utility classes; conflicts with sibling PRs resolve by
re-running the script.
Nothing is ever written to pack files — canonical values live only in
the song_enrichment display cache. Cover art caching and acoustic
fingerprinting are follow-up slices.
22 pure unit tests + 19 server tests (fake transport injected over the
_mb_http_get seam) + demo-mode route cases; full-suite failure set
A/B-identical with the change stashed vs applied.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Nm7tHs1Yvjjtnnu4nzJgdN
* library: match-review modal + configurable auto-apply confidence (P8 R0)
Follow-up to the initial P8 commit, folding in the first round of tester
feedback on the review surface and the matcher's knobs:
- Review GUI is a centred MODAL now, not a sidebar — one chart at a
time (the scraper-review model from media-server / emulation-frontend
apps): the chart's current metadata with explicit amber
"Missing: album / year / cover art" chips (art detected via the art
request failing), candidates each carrying "Adds: year - genres -
ISRC" / "Shows as: ACDC -> AC/DC" per-field chips, and Skip /
Not a match / Search instead / Use selected with prev-next + arrow-key
navigation. Chip + window API surface unchanged, so songs.js needed no
edits for the rework.
- Auto-apply confidence is a SETTING: default drops 0.95 -> 0.90
(mb_match.AUTO_MIN; classify() takes an auto_min override). The
per-field floors are untouched and threshold-independent — a
perfect-title cover by the wrong artist still can't auto-match at any
setting. New validated settings keys: enrich_enabled (bool) +
enrich_auto_threshold (0.5–1.01; >1.0 = "Always review", since a
capped score can equal exactly 1.0). Read once per pass; disabling
gates only the BACKGROUND matcher — manual search/fix stays available.
- Settings -> Library -> "Metadata matching" card: enable toggle,
confidence select (85 / 90 / 95 / Always review), a Match Now button
(new POST /api/enrichment/kick, single-flight like every other kick,
demo-mode blocked), and a live status line fed by the same fetch as
the review chip. Markup in index.html per the v3 settings pattern,
wired by match-review.js, null-guarded so v2 no-ops.
- Review queue orders missing-data charts first — confirming those has
the most to gain; complete charts only stand to be re-labelled.
Tests: threshold moves the auto/review boundary via settings; the
enable toggle gates matching but not the manual proxy; settings
validation; kick route; queue ordering; classify(auto_min=...) floors.
Full-suite failure set byte-identical to the pre-change baseline.
tailwind.min.css regenerated for the modal's utility classes.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Nm7tHs1Yvjjtnnu4nzJgdN
* fix(library): lock MusicBrainz throttle across sleep + de-dup enrich queue (PR #710 review)
Hold a module-level lock across _enrich_throttle's read/sleep/write so the
background daemon and threadpooled sync search route serialize outbound MB
requests instead of bursting past the 1 req/s limit. De-dup the enrich queue
by filename so a changed-hash failed row isn't processed twice per pass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: byrongamatos <xasiklas@gmail.com>
|