Ablation Study
6 configsRetrieval quality benchmark with a fixed query/chunk corpus and precomputed embeddings.
CodeSearchNet (Python annotated + JS/Java/Go/PHP/Ruby HuggingFace) — 249 queries — 2848 candidates — human-labeledembedding: sentence-transformers/all-MiniLM-L6-v2 (precomputed)Resultshover row for description
| Config | Recall@5 | MRR | NDCG@10 | Latency |
|---|---|---|---|---|
| Full Pipelinebaseline | 62.9% | 0.766 | 0.710 | 38533μs |
| No Quantization | 59.8% | 0.765 | 0.709 | 24245μs |
| Vector-Only | 57.0% | 0.733 | 0.670 | 13752μs |
| No Reranking | 61.4% | 0.759 | 0.704 | 24236μs |
| CodeRAG Multi-Path | 57.8% | 0.715 | 0.671 | 69238μs |
| PageIndex (Keyword) | 0.1% | 0.006 | 0.002 | 23758μs |