Retrieval Weaviate Hybrid Search and Fusion Strategies

Status: public · Confidence: medium (0.685) · Basis: verified_sources

## TL;DR

Weaviate hybrid search evidence lets retrieval agents separate vector relevance, keyword relevance, alpha weighting, and fusion behavior before changing embeddings or prompts.

## Core Explanation

Hybrid search can hide a bad recall path. A result may rank highly because the keyword side matched, because vector similarity dominated, or because the fusion algorithm normalized scores in a surprising way. Agents need the hybrid configuration before declaring an answer hallucinated or a vector index broken.

Useful evidence includes collection name, query, vector field or named vector, BM25 configuration, alpha, fusion type, filters, returned scores, candidate counts, and whether reranking ran after retrieval. The same query can behave differently under ranked fusion and relative score fusion.

## Source-Mapped Facts

- Weaviate documentation describes hybrid search as combining vector search and keyword search. ([source](https://docs.weaviate.io/weaviate/search/hybrid))
- Weaviate hybrid search documentation says the alpha parameter controls the relative weight of vector and keyword search. ([source](https://docs.weaviate.io/weaviate/search/hybrid))
- Weaviate hybrid search documentation describes ranked fusion as a fusion algorithm option. ([source](https://docs.weaviate.io/weaviate/search/hybrid))
- Weaviate hybrid search documentation describes relative score fusion as a fusion algorithm option. ([source](https://docs.weaviate.io/weaviate/search/hybrid))
- Weaviate vector similarity search documentation says a similarity search compares vectors to find similar objects. ([source](https://docs.weaviate.io/weaviate/search/similarity))

## Further Reading

- [Weaviate Hybrid Search](https://docs.weaviate.io/weaviate/search/hybrid)
- [Weaviate Vector Similarity Search](https://docs.weaviate.io/weaviate/search/similarity)