Hybrid search — vector + full-text in one query.
Reciprocal Rank Fusion (RRF) combines vector similarity scores and BM25 rankings into a single relevance ranking. Pair with metadata filters for precision and recall you can't get from either approach alone.
Vector alone
Good at semantic similarity
Vector search captures meaning and concept. It finds "automobile" when you search "car" — but it can miss exact keyword matches for product codes, names, and jargon.
Full-text alone
Good at keyword precision
BM25 excels at exact keyword retrieval — SKU numbers, names, acronyms. But it misses paraphrases and semantic context, returning nothing for synonyms.
Hybrid (RRF)
The best of both worlds
RRF fusion ranks candidates by their combined position in both result lists. Documents that score well on both signals float to the top. Precision and recall together.
RRF fusion with full control.
Fusion
Reciprocal Rank Fusion
RRF combines ranked lists without normalizing scores across different scales. Score(d) = Σ 1/(k + rank(d)) where k=60. Robust and parameter-light.
Weighting
Adjustable signal weights
Control how much each signal contributes with alpha weighting. alpha=1.0 is pure vector; alpha=0.0 is pure full-text; alpha=0.5 is balanced hybrid.
Filters
Pre-filter before fusion
Apply metadata filters before the hybrid search runs. Only candidates matching the filter are scored — reducing noise and improving precision for structured corpora.
The best of both retrieval worlds.
RRF
fusion algorithm
2×
higher recall vs single-mode
<15ms
p50 hybrid query
1
API call for both modes
Common questions
What is RRF and why does ManyVector use it?
Reciprocal Rank Fusion merges ranked lists from vector and full-text search. It is parameter-free, robust to score-scale differences, and consistently outperforms weighted-sum approaches in benchmarks.
Do I make two separate API calls for hybrid results?
No. Send one query() call with both vector and query_text. ManyVector handles parallel retrieval and fusion server-side and returns a single ranked list.
How much better is hybrid vs vector-only search?
In benchmarks across RAG corpora, hybrid search improves recall@5 by 15-30% vs vector-only — especially for queries with specific entity names, product codes, or exact phrases.
Can I weight vector results more than full-text?
RRF itself is weight-free by design. You can bias results further by combining filter conditions with your hybrid query to promote or demote specific attribute values.