Retrieval is the part of a question-answering system that nobody sees and everybody feels. If the right passages never reach the model, no amount of prompting saves the answer. So the question worth asking is narrower than “is our search good?” It’s: “did the most relevant context arrive first?”
Two Ways to Search, Two Different Blind Spots
Semantic search compares the meaning of a question to the meaning of your documents. It is good at finding passages that say the same thing in different words. Ask about “termination clauses” and it will surface text about ending an agreement even if those exact words never appear.
Keyword search does the opposite. It matches the literal terms in your query. It is precise about names, codes, and rare phrases, but it cannot connect ideas that are worded differently.
- Semantic search misses exact terms: part numbers, acronyms, statute references, names.
- Keyword search misses meaning: paraphrases, synonyms, and indirect phrasing.
- Most real questions need both at once.
Key Insight
Choosing semantic-only or keyword-only is choosing which half of your questions to answer badly. Hybrid search refuses the trade-off.
What Hybrid Search Actually Does
Hybrid search runs both methods and then merges their results into a single ranked list. Each method produces its own ordering of candidate passages. A fusion step combines those orderings so that a passage ranked highly by either method, and especially by both, rises to the top.
The practical effect is forgiving retrieval. A query that leans on exact wording still benefits from meaning, and a query that leans on meaning still benefits from exact matches. You stop tuning your system around the shape of a single question.
Why You Still Need Reranking
Hybrid search gets the right passages into the candidate set. It does not guarantee the single best passage sits at the very top. Fusion is fast and approximate by design; it ranks using signals computed before the question and the passage are read together.
A reranker reads each candidate against the actual question and scores how well it answers it. It is slower, so you only apply it to the handful of passages hybrid search already surfaced. The result is a final ordering that reflects relevance to this specific question, not general similarity.
- 1.Hybrid search casts a wide, balanced net and returns a short list of candidates.
- 2.A reranker reads that short list against the question and reorders it.
- 3.Only the top results, now genuinely the most relevant, are handed to the model.
Bring Your Own Reranking
Reranking quality depends on the model doing the scoring, and the right model differs by domain and language. Noesia treats reranking as something you configure rather than something hard-wired. You point a collection at the reranking model you trust, with your own credentials, and Noesia applies it to the candidates from hybrid search.
If reranking is not configured, search still returns sensible results from hybrid retrieval alone, and the response says so plainly. Nothing fails silently. You always know whether reranking was applied, skipped, or unavailable.
Better answers rarely come from a bigger model. They come from making sure the right context arrives first. Hybrid search widens what you can find; reranking decides what the model actually reads.
The model can only reason about what retrieval puts in front of it.