Fix source-missing auto-merge and remove Pink Floyd examples from prompts
All checks were successful
Publish Metadata Agent Image (dev) / build-and-push-image (push) Successful in 1m10s
Publish Web Player Image (dev) / build-and-push-image (push) Successful in 1m10s

Auto-merge: when ingest pipeline detects "source file missing", now checks
if the track already exists in the library by file_hash. If so, marks the
pending entry as 'merged' instead of 'error' — avoiding stale error entries
for files that were already successfully ingested in a previous run.

Prompts: replaced Pink Floyd/The Wall/Have a Cigar examples in both
normalize.txt and merge.txt with Deep Purple examples. The LLM was using
these famous artist/album/track names as fallback output when raw metadata
was empty or ambiguous, causing hallucinated metadata like
"artist: Pink Floyd, title: Have a Cigar" for completely unrelated tracks.

Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-03-20 01:05:22 +00:00
parent 8d70a5133a
commit 71d5a38f21
3 changed files with 28 additions and 16 deletions

View File

@@ -4,14 +4,14 @@ You are a music library artist merge assistant. You will receive a list of artis
You will receive a structured list like:
### Artist ID 42: "pink floyd"
Album ID 10: "the wall" (1979)
- 01. "In the Flesh?" [track_id=100]
- 02. "The Thin Ice" [track_id=101]
### Artist ID 42: "deep purple"
Album ID 10: "machine head" (1972)
- 01. "Highway Star" [track_id=100]
- 02. "Maybe I'm a Leo" [track_id=101]
### Artist ID 43: "Pink Floyd"
Album ID 11: "Wish You Were Here" (1975)
- 01. "Shine On You Crazy Diamond (Parts I-V)" [track_id=200]
### Artist ID 43: "Deep Purple"
Album ID 11: "Burn" (1974)
- 01. "Burn" [track_id=200]
## Your task
@@ -20,7 +20,7 @@ Determine if the artists are duplicates and produce a merge plan.
## Rules
### 1. Canonical artist name
- Use correct capitalization and canonical spelling (e.g., "pink floyd" → "Pink Floyd", "AC DC" → "AC/DC").
- Use correct capitalization and canonical spelling (e.g., "deep purple" → "Deep Purple", "AC DC" → "AC/DC").
- If the database already contains an artist with a well-formed name, prefer that exact form.
- If one artist has clearly more tracks or albums, their name spelling may be more authoritative.
- Fix obvious typos or casing errors.
@@ -54,7 +54,7 @@ Determine if the artists are duplicates and produce a merge plan.
You MUST respond with a single JSON object, no markdown fences, no extra text:
{"canonical_artist_name": "...", "winner_artist_id": 42, "album_mappings": [{"source_album_id": 10, "canonical_name": "The Wall", "merge_into_album_id": null}, {"source_album_id": 11, "canonical_name": "Wish You Were Here", "merge_into_album_id": null}], "notes": "..."}
{"canonical_artist_name": "...", "winner_artist_id": 42, "album_mappings": [{"source_album_id": 10, "canonical_name": "Machine Head", "merge_into_album_id": null}, {"source_album_id": 11, "canonical_name": "Burn", "merge_into_album_id": null}], "notes": "..."}
- `canonical_artist_name`: the single correct name for this artist after merging.
- `winner_artist_id`: the integer ID of the artist whose record survives (must be one of the IDs provided).