Semantic search understands that “make my app faster” relates to performance, but it is poor at exact strings — error codes, function names, product names, version numbers — where a slight embedding difference buries the exact match. Keyword search is the reverse: precise on literals, blind to paraphrase.

Running both and fusing the rankings covers each one's blind spot. It is why a search that returns nothing for a specific tool name and everything for a vague description usually needs a second retrieval path, not a better model.