Why does Weaviate combine BM25 and vector search instead of using vector search alone?

Dense vector embeddings excel at capturing semantic meaning and paraphrased similarity, but they can genuinely struggle with specific, rare terms, an exact product code, an uncommon proper noun, a precise legal citation, where the literal token matters and a general-purpose embedding model may not have encoded that specific term distinctly enough to surface it reliably. BM25's term-frequency-based scoring, by contrast, is very good at exactly this kind of literal, exact-match relevance, but has no notion of semantic meaning or paraphrase at all.

Combining both signals means a query benefits from each approach's strengths without inheriting either one's specific weakness: a search for an exact product SKU still surfaces the right result even if the semantic embedding alone might not have ranked it highly, while a paraphrased, conversational query still benefits from the vector component's ability to match meaning rather than requiring literal term overlap.

This is precisely why independent benchmarking on domain-specific corpora, technical documentation, legal retrieval, code search, tends to show hybrid search outperforming either pure vector or pure keyword search alone: these are exactly the domains where both exact terminology and semantic understanding matter simultaneously, rather than one clearly dominating the other.