Fusing BM25 and dense rankings

Nineteen Larkspur Cloud handbook chunks, ranked two ways: real BM25, and cosine similarity from the all-MiniLM-L6-v2 embedding model. Pick a question, pick how to fuse the two rankings, and move alpha. The three chunks above the magenta rule are what the LLM receives.

Golden question

Fusion method

Weighted scores Reciprocal rank fusion

Weight on dense, alpha

RRF constant c

Chunks each retriever hands over

4, the LangChain default 10

Evidence rank, BM25
Evidence rank, dense
Evidence rank, fused
Recall@3, fused

Fused ranking for this question

Rank	Chunk	BM25	Dense	Fused	Golden set

Small figures are each retriever’s own rank. BM25 hands over only chunks that share a term with the question. Ties keep handbook order.

Recall@3 on the golden set

Golden question	BM25	Dense	Fused
Mean recall@3
Questions with all evidence in the top 3

BM25 and dense columns rank with one retriever alone. The fused column follows the controls; magenta marks a question only the fused list answers.