Hybrid search runs two kinds of retrieval on the same query, usually a keyword method such as BM25 and a vector method based on embeddings, then merges the results into one list. Each method covers the other's weak spots: keywords catch exact terms, while vectors catch paraphrases and related concepts. Hybrid retrieval is a common pattern in retrieval-augmented generation systems, though specific AI search products rarely say exactly how they combine signals.
The two halves
Keyword retrieval scores documents on how often the query terms appear and how rare those terms are across the collection. BM25 is the standard formula. It is fast, predictable and very good at exact matches.
Vector retrieval compares the meaning of the query with the meaning of each passage. It finds relevant passages that share few words with the query.
Merging the results
Because the two methods produce scores on different scales, systems often merge by rank rather than raw score. Reciprocal rank fusion is a common, simple approach: each document gets credit based on its position in each list, and the totals decide the final order. A reranking model may then reorder the top results.
RRF score(d) = Σ 1 / (k + rank_i(d))
// summed over each result list i
// k is a constant, often set around 60Implications for content
If a system uses both signals, a page benefits from containing the exact terms people search for and from explaining the concept clearly. In practice that means stating the product name, feature name and category term in the passage, and then explaining what they mean in plain language.
Example
A help center assistant
Cobaltreach, a fictional email deliverability tool, builds an assistant on its help center. Users ask both 'why are my emails going to spam' and 'what does error 550 5.7.1 mean'. Vector retrieval handles the first well; keyword retrieval handles the second. With hybrid retrieval, both kinds of question find the right article.
Common mistakes
Adding raw scores together
Keyword and vector scores are not on comparable scales. Rank-based fusion or a trained reranker is the usual fix.
Assuming one product's design applies to all
Commercial AI search systems do not publish their retrieval stacks. Use hybrid search as a model for your thinking, not as a confirmed description of any one engine.
Questions
A long-established keyword ranking function that scores documents by term frequency, term rarity and document length. It remains a strong baseline in information retrieval.
For docs and help centers with many exact terms, such as error codes and API names, hybrid retrieval is often worth it. Test on real queries before deciding.
