andgineer / news-recap / 37474585299

88%

Build Type

push

github

Committed by andgineer

Commit Message

Dedup on Jev

Deduplicate asks Jev "same piece of news?" for every pair inside the embedding
candidate groups (merge at >= 0.40) and for a wider net of pairs at similarity
0.85-0.90 that today's pipeline never examines (merge at >= 0.70). Merge groups
are star-grouped components; a merged group keeps an existing headline.
NEWS_RECAP_DEDUP_BACKEND=llm|jev, default jev when a key is found; on Jev
failure the step falls back to the LLM.

Gate (bench_jev.py dedup, labels in spec/plan-jev-decisions.md Stage 5):
wrong merges 13 vs 31 for today's LLM on the tuning nights, 1 vs 2 on the
holdout, with as many correct pair decisions; the wider net is 92% right.
agy -6 launches / -157k tokens on 10-06; Jev ~$1.5-2/month.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Coverage Stats

160 of 163 new or added lines in 4 files covered. (98.16%)

29 existing lines in 3 files now uncovered.

5488 of 6374 relevant lines covered (86.1%)

0.86 hits per line