Start with the query, not the index Keyword search finds exact terms, product codes, and error messages well. Vector search can retrieve passages whose mea...
Start with the query, not the index
Keyword search finds exact terms, product codes, and error messages well. Vector search can retrieve passages whose meaning resembles a question even when they use different words. Neither is universally better. A support knowledge base may need both: an exact error identifier and a natural language description of the same problem.
Build a set of real user queries and label the documents that answer them. Include short identifiers, long questions, typos, synonyms, and requests that have no answer. Compare retrieval quality on that set before choosing infrastructure.
How keyword retrieval works
A text index tokenizes content and ranks matches based on terms and fields. It is predictable for exact names and can highlight why a document matched. Synonyms and stemming improve recall, but aggressive expansion can introduce irrelevant matches. Tune title, heading, and body weighting with actual queries.
How vector retrieval works
An embedding model maps a query and passages into vectors. A similarity search returns nearby passages, which can help when the question paraphrases the source. It can also retrieve thematically related but factually wrong passages. Model choice, chunk boundaries, and index freshness matter.
Do not treat a vector score as a calibrated probability that the answer is correct. It is a ranking signal inside a particular model and corpus.
Combine and rerank
Hybrid retrieval gathers candidates from both methods. Merge and deduplicate them, then optionally rerank the candidate set. The extra stage can improve precision but adds latency and cost. Measure whether it helps the questions users actually ask.
Apply tenant and document permissions before candidates reach the answer model. If either retrieval branch bypasses the access filter, the hybrid system is unsafe. Store source version and location so an answer can cite the right passage.
Evaluate the failure modes
Measure whether a known relevant source appears in the top few results and inspect misses by query class. Exact identifiers missed by vectors suggest a keyword branch. Conceptual questions missed by keyword search suggest semantic retrieval. Check "no answer" queries to see whether the system returns plausible but unrelated passages.
Conclusion
Choose retrieval from observed query behavior, not a preference for newer technology. An assistant needs relevant, current, permission checked passages before generation can be trusted. See AI and business automation for the application context.
References