Skip to content

feat: add Jina neural reranker — LoCoMo QA 53.2% → 57.3%#61

Merged
CodeNinjaSarthak merged 4 commits into
mainfrom
dev
Apr 9, 2026
Merged

feat: add Jina neural reranker — LoCoMo QA 53.2% → 57.3%#61
CodeNinjaSarthak merged 4 commits into
mainfrom
dev

Conversation

@CodeNinjaSarthak

Copy link
Copy Markdown
Owner

Summary

  • Integrate Jina neural reranker into retrieval pipeline for post-semantic-search reranking
  • Benchmark on full LoCoMo dataset (n=1540, all 10 conversations) — overall QA accuracy improved from 53.2% to 57.3% (3.87x over baseline)
  • Category breakdown: Temporal 67.3%, Open-domain 62.0%, Single-hop 38.7%, Multi-hop 38.5%

Test plan

  • make lint passes
  • make test passes
  • Verify reranker integrates correctly with existing retrieval pipeline

- Add checkpoint/resume + parallelism + retry to ingestion script
- Add LLMError and content filter handling to retryable errors
- Add 60s timeout to Azure client to prevent silent hangs
- Add eval_qa_rag_baseline.py with resume support
- Add concurrency=5 + resume + content filter to eval_qa_accuracy.py
- Add 10 regression tests for _parse_batch_result
- Update .gitignore to exclude eval/checkpoints/

Results: pipeline 46.6%, RAG baseline 44.4% across all 10 LoCoMo convs
Temporal gain: +31.5pp (pipeline vs RAG)
Results (n=1540, all 10 LoCoMo convs):
- Pipeline v2: 46.6% overall (+2.2pp over RAG baseline 44.4%)
- Temporal: +32.4pp over RAG (57.3% vs 24.9%) — dominant gain
- Open-domain regression confirmed architectural, not two-pass related

Changes:
- Add per-category two-pass gate (category != 4) to eval_qa_accuracy.py
- Add two_pass_used field to per-pair results for paper analysis
- Update result.md with final all-10 benchmark numbers and SOTA comparison

Next: BM25 + RRF, cross-encoder reranker, temporal date filtering
  - Integrate Jina reranker for neural reranking after semantic search
  - Update LoCoMo benchmark results: 53.2% → 57.3% overall QA accuracy (n=1540, all 10 convs)
  - Update uv.lock with new dependencies
  - Add cl.md to .gitignore
@CodeNinjaSarthak CodeNinjaSarthak merged commit ccb2fe1 into main Apr 9, 2026
2 checks passed
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