- Dedup: mark only accepted picks as seen (not all prefiltered
candidates) — unselected items stay eligible for re-evaluation,
preventing pool exhaustion across runs
- Queries: expanded from 29 to 37+ with rotating 30-subgenre list,
25 label targets, 14 review sites; Bandcamp/MA queries skip
time_range for broader results; review sites use time_range:year
- Results per query: 15 → 25
- Prefilter: parallel batches of 35 (up to 3 concurrent), processes
all fresh candidates instead of just top 80; be-inclusive prompt
- Curator: cap 20 → 30, score floor 60 → 50, URL prefix matching
in provenance check instead of exact match
Result: 405 candidates/run vs 146 before; 88 passing prefilter vs 10;
pool stays at ~400 fresh on consecutive runs.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
FastAPI service replacing the 77-node n8n pipeline. Implements full
discovery pipeline: 29 rotating SearXNG queries, nomic-embed-text
scoring against Last.fm taste centroid, Mistral-nemo prefilter and
curator with provenance validation, SQLite dedup, writes to
metal-picks.json for the existing FGS dashboard.
Runs as systemd service on port 8766 (fgs-agent.home via Caddy).
n8n reduced to a 2-node schedule trigger → HTTP POST.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
nick2day