01 / Featured investigationRAG & Evaluation
October 20268–10 min read
Why Adding BM25 Made My RAG Retrieval Worse
I expected adding BM25 to improve my RAG retrieval pipeline. Instead, hybrid search performed worse. Here’s how systematic evaluation led to query-aware retrieval.
THE QUESTIONWhen does an extra retrieval signal help—and when does it get in the way?
Read the investigationTHE RETRIEVAL EXPERIMENT
More signals.
Better retrieval?
Mean reciprocal rank0 → 1
- DenseSemantic baseline
- 0.743
- Naive HybridUnconditional fusion
- 0.705
- Query-Aware HybridConditional hybrid
- 0.883
Development results, not an independent production evaluation.