Semantic search handles paraphrase well, but struggles with identifiers, names, acronyms and exact phrases. Keyword search is the reverse. Combining them covers both.
Where each is stronger
- Vector — conceptual questions, differently worded queries, natural language.
- Keyword — part numbers, error codes, legal references, names and rare terms.
How to combine
- Run both searches and merge the results, with a ranking method you can tune.
- Apply filters — date, document type, permissions — before ranking.
- Evaluate on your own test set, including queries with exact identifiers.
If users complain that search 'can't find the thing I know exists', check whether exact-match queries are being handled.