Skip to main content
Retrieval is what happens when your Agent needs to look something up. Given a query and a data source, retrieval returns the small set of chunks most likely to be useful. Good retrieval is often the difference between an Agent that answers well and one that hallucinates. Nora ships strong defaults but every knob is exposed.

​
The pipeline, in order

  1. Query construction — the Agent’s question, possibly rewritten for clarity.
  2. Filter — narrow candidates by metadata (tag, source, date, entity).
  3. Rank — hybrid vector + keyword scoring.
  4. Fuse — combine signals into a single relevance score.
  5. Ground (optional) — check against a causal graph before returning.
  6. Cut off — return the top-N chunks by score.
Each step is configurable via a Retrieval preset.

​
Where retrieval happens

  • Inside an Agent block — every model call that has attached data sources triggers retrieval first.
  • Inside a Tool — a special retrieve tool lets the Agent explicitly search when it decides it needs to.
  • In a Foundry pipeline — for chained pipelines where one output feeds another’s context.

​
What you can tune

Hybrid search

Balance vector similarity vs. keyword matching.

Presets

Save a retrieval config and reuse across Agents.

Filters

Metadata filters — tag, source, entity, date.

Ranking

Weights, boosts, and reranker toggles.

Grounding

Verify retrieved facts before returning.

Diagnostics

Why isn’t my chunk showing up?