Skip to main content
Nora’s retrieval defaults to hybrid search: two independent ranking passes (vector and keyword) fused into one final score. This works better than either alone.

​
Why hybrid

  • Vector search finds chunks that mean similar things, even if wording differs. Great for paraphrased questions, cross-language, and conceptual matches. Weak on rare exact terms — SKUs, error codes, proper nouns.
  • Keyword (BM25) search finds chunks with the same literal terms. Great for identifiers, jargon, exact-string lookups. Weak on synonyms and paraphrases.
Most real questions need both.

​
How fusion works

Vector and keyword each produce a ranked list. Nora combines them using Reciprocal Rank Fusion (RRF) — a chunk in position 3 of both lists scores higher than a chunk that’s #1 in one list and #100 in the other. RRF is robust to score scale differences and has no tunable magic constant. The fused list is what the Agent sees.

​
Adjusting the balance

Even with RRF, you can weight the passes. From your Retrieval preset:
  • Vector weight — 0.0 to 1.0. Default 0.5.
  • Keyword weight — 0.0 to 1.0. Default 0.5.
Push toward vector for conversational domains (support, legal Q&A). Push toward keyword for code, part numbers, product names. Setting one weight to 0 turns off that pass entirely (pure vector or pure keyword).

​
Query expansion

For hard queries — short, ambiguous, or single-keyword — Nora can rewrite the query into a small set of variants and search each. Results are re-fused. Off by default because it adds latency and cost. Turn on in your preset: Query expansion → on. Set expansion count (2-5). Uses a small model.

​
Multilingual

Vector search works across languages when your embedding model is multilingual. Keyword search matches only the surface language. If your data is multilingual, embed with a multilingual model and lean the weight toward vector.

​
Testing

Every Retrieval preset has a Test query panel. Type a query, see the ranked results, and toggle vector-only / keyword-only / hybrid to compare. Score breakdown is shown for every returned chunk.

​
When hybrid is overkill

For very small collections (< 1000 chunks) keyword-only is often enough and costs nothing extra. For very short, keyword-heavy queries (SKU lookups) keyword-only is faster with the same recall. Turn off vector in those presets.