Different queries need different signals
An exact standard identifier is a strong lexical signal. A conceptual description of a technical mechanism may need semantic retrieval. A single search experience often needs to support both.
Build separate candidate sets
Use the keyword engine to retrieve exact terms and field matches. Use an embedding index to retrieve conceptual neighbors. Preserve the metadata needed for filtering and source attribution in both paths.
Merge, then rerank
Rank fusion combines differently scored candidate lists without pretending that their raw scores are equivalent. A reranker can then examine a smaller candidate set with more query context.
Measure relevance on your own corpus
Create representative query sets and document expected results. Include terminology differences, exact identifiers and intentionally unanswerable queries. Evaluate latency and relevance together.
An architectural perspective, illustrated through general engineering patterns. Implementation choices depend on your data, permissions and operational constraints.