Hybrid Search — AI Glossary

Reference details

Related terms

• BM25
• Semantic Search
• Reranker (Cross-Encoder)
• Reciprocal Rank Fusion (RRF)
• Context Recall

How it works

1. Two retrievers run over the same corpus and return their own ranked lists, which cannot be combined by score because a cosine similarity and a BM25 score are not on a common scale.
2. Reciprocal Rank Fusion solves this by discarding the scores entirely and using only the position: each document scores one over a constant plus its rank in each list, and the sums are added.
3. With the usual constant of 60, a document ranked first in one list and tenth in the other scores 1/61 + 1/70 ≈ 0.0307, while a document ranked third in both scores 1/63 + 1/63 ≈ 0.0317.
4. So consistent mid-rank agreement outranks one strong placement paired with a weak one, which is the behaviour the method is chosen for — corroboration across retrievers is treated as evidence.
5. The constant controls how sharply early ranks dominate: a smaller value weights the top positions more heavily, and raising it flattens the contribution of rank differences.
6. Because fusion reads only ranks, either retriever can be replaced or retuned without renormalising anything — and equally, a retriever returning plausible-looking rubbish at rank one is given full weight by construction.

Example: The query each half gets wrong

Search “ORA-01555 snapshot too old”. Dense retrieval embeds it near general database-error text and may return the wrong code. BM25 matches the literal string and nails it. Now search “why does my long-running report keep failing” — BM25 finds nothing useful because the document never uses those words, and dense retrieval finds it immediately. Hybrid runs both and fuses the rankings, which is why production retrieval rarely picks a side.

Commonly confused with

Hybrid search is not simply running two searches and concatenating results. The fusion step is the substance: two rankings with incomparable score scales have to be combined into one order. Reciprocal Rank Fusion is common precisely because it uses rank position and ignores the raw scores, which are not on the same scale.

When to use it

Reach for it when:

• Queries mix natural language with identifiers, codes or product names
• Either method alone shows recall gaps you can point at in evaluation
• You can afford two index lookups per query

Reach for something else when:

• Queries are uniformly one kind — all lexical or all conversational
• You have not measured which half is failing; hybrid can mask a fixable retriever
• Latency budget will not absorb a second search plus fusion

Referenced by

This term is used by the following ModelRefs references:

• learn rag
• build enterprise search

Where this appears on ModelRefs

• BGE-M3
• measure retrieval recall and grounding

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Hybrid Search — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path