Recherche hybride et fusion de rangs
Querying two engines that fail in different ways, then merging their rankings. Vector search finds meaning and misses the exact reference; keyword search (BM25) finds the literal term and misses the paraphrase. Their scores live on incompatible scales, so averaging them is meaningless. Reciprocal rank fusion sidesteps this by looking only at positions: each document gets the sum, across every list, of 1 divided by (k plus its rank), with k defaulting to 60. That output is a first pass, before a reranking model sorts the short list handed to the generator.
Strengths
- No tuning to find: fusion works on ranks, so it holds even between two unrelated engines
- Fixes the most visible failure of a home-grown RAG, the exact reference a user hunts for and never finds
- Already implemented natively by mainstream search engines, so there is rarely code to write
Limitations
- Two engines to feed and keep in sync, which makes indexing heavier to operate
- The value 60 is a default inherited from very large corpora, and deserves questioning on a base of a few hundred pages
- Does not rescue bad chunking: if the useful piece was cut in half, no fusion will glue it back
Best for
- A corpus mixing prose and identifiers, product catalogues, technical documentation, legal bases
- Diagnosing a RAG that answers off-target before swapping model or vendor