Switch reranker from Jina API to local cross-encoder#64
Merged
Conversation
- Archive 18 superseded/debug result files (untracked) - Rename canonical result files for clarity - Add --no-isolation and --no-rerank flags to eval script - Add 5-attempt retry on generation, retrieval, judge calls - Add run_ablation_parallel.sh for parallel ablation runs - Record no_isolation/no_rerank/model in result metadata
Update README numbers to 56.3% overall, 64.2% temporal (local cross-encoder/ms-marco-MiniLM-L-6-v2 results, n=1540) Replace Jina branding in mermaid diagrams and text Add ablation rows showing reranker is load-bearing component Fix fact extraction precision discrepancy (58.6% -> 52%) Remove JINA_API_KEY from .env.development.example Update eval_qa_accuracy.py help strings for --no-rerank flags Document --local-rerank flag in Reproduce Results section
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
cross-encoder/ms-marco-MiniLM-L-6-v2) loaded viasentence-transformers. The local model runs on CPU with ONNX INT8 quantization and produces results within noise of the Jina API while removing all API key, rate limit, and reproducibility concerns.Results (LoCoMo, n=1540)
Key finding
Ablation reveals the cross-encoder reranker is the load-bearing component. Round-robin merge vs. score-based merge produces identical results (55.8%) once the reranker is in place — the merge strategy becomes irrelevant. Without the reranker, neither merge strategy improves over isolation alone.
Changes
README.md: numbers updated, Jina references replaced, 4 new ablation rows added, fact extraction precision corrected (58.6% → 52% to match measured value).env.development.example:JINA_API_KEYsection removedeval/eval_qa_accuracy.py:--local-rerank,--merge-strategy,--top-k,--fetch-multiplier,--no-rr-rerankflags added; help strings updatedNote:
JinaRateLimiterdead code (424f359) is intentionally left in — cleanup belongs in a follow-up PR.Test plan
--local-rerankflag on a single conv before full runJINA_API_KEYis no longer referenced in.env.development.example--no-rerankablation still works (no reranker path)