Skip to content

perf(retrieval): tune hybrid recall parameters - #66

Merged
juemimgcd merged 1 commit into
masterfrom
codex/rag-recall-tuning
Sep 1, 2026
Merged

juemimgcd merged 1 commit into
masterfrom
codex/rag-recall-tuning

Conversation

@juemimgcd

Copy link
Copy Markdown
Owner

Summary

  • expand each retrieval backend candidate pool from 50 to 100
  • tune the lexical RRF weight from 0.30 to 0.55
  • preserve the existing Top-10 output contract and scoped retrieval filters

Evaluation

BEIR SciFact, 300 test queries, 4,000 documents, Recall@10:

  • Recall: 0.812222 -> 0.822889 (+0.010667)
  • MRR: 0.631085 -> 0.634030 (+0.002945)
  • NDCG: 0.668994 -> 0.673534 (+0.004540)

Validation

  • Ruff: passed for the changed file
  • Existing scoped retrieval tests: 3 passed
  • Retrieval package compileall: passed
  • git diff --check: passed

No test files were modified.

@juemimgcd
juemimgcd merged commit 61f4fe9 into master Sep 1, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant