An engineer on the platform team flagged the internal knowledge assistant for returning a three‑paragraph answer about exponential backoff when she asked for the custom “dead‑letter queue threshold” document. The relevant file sat just outside the top‑ten results, meaning dense‑only retrieval missed it. The incident exposed the limits of pure semantic search in production RAG pipelines, in a high‑stakes environment where incorrect guidance can cause outages.

Before neural vectors, BM25 dominated keyword search, scoring rarity, frequency and length to surface exact matches. Because it ignores synonyms, it complements dense embeddings that excel at conceptual queries. Weaviate’s hybrid mode blends BM25 and vector scores via a tunable alpha parameter; an alpha of 0.5 gives equal weight to both signals, shifting results toward documents containing the precise term.

Running a 150‑query evaluation on the engineering corpus showed pure BM25 hit‑rate 0.71, pure dense 0.73, while a 0.5 alpha achieved 0.83 hit‑rate and 0.69 MRR, moving the missed “dead‑letter queue” record from rank eleven to rank four. The experiment demonstrates that a modest hybrid blend consistently outperforms either extreme for mixed conceptual and exact‑term searches.