Skip to main content
Retrieval brings back the most relevant chunks — but “relevant” isn’t “true”. Grounding is an optional pass that cross-checks retrieved content against your causal knowledge graph before it reaches the Agent. Grounding catches:
  • Chunks that are relevant but out of date (superseded facts).
  • Chunks that contradict facts you’ve established as authoritative.
  • Chunks that fill in around a fact but don’t actually contain it.

​
When grounding helps

  • Regulated content — legal, medical, financial answers where being off matters.
  • Fast-moving domains — pricing, availability, policy, hours.
  • Numeric answers — where a wrong number is worse than “I don’t know.”
For casual Q&A, grounding is overhead. Turn it off.

​
How it works

  1. Retrieval returns candidate chunks.
  2. Each chunk is parsed for factual claims — dates, quantities, entity relationships.
  3. Claims are checked against the causal graph attached to the retrieval preset.
  4. Chunks that contradict the graph are demoted or dropped.
  5. Chunks the graph confirms get a verified tag the Agent can cite.

​
Setup

  1. Build a causal graph — see Causal Graph overview.
  2. Attach it to your Retrieval preset — Grounding → Causal graph → pick.
  3. Choose action on conflict:
    • Drop — remove contradicting chunks.
    • Demote — score them 0.5× and keep.
    • Flag — return them with a warning field the Agent can surface.

​
Confidence

Not every chunk contains a checkable claim. Grounding runs a fast pre-check and only invokes the graph verifier on chunks that look verifiable. Uncheckable chunks pass through untouched. The Trace view shows for each chunk: grounded / demoted / dropped / unchecked.

​
Bidirectional benefit

Grounding also flags missing claims — cases where the graph has an authoritative fact that retrieval failed to surface at all. This surfaces as an Improvement candidate in Signals (see Improvement proposals).

​
When to skip grounding

  • You don’t have a causal graph. Set one up first — see Extract from documents.
  • Your data is opinion or unstructured commentary where “correctness” doesn’t apply.
  • You need every millisecond of latency (grounding adds 50-200ms per query).

​
Debugging

If Agents complain “the docs say X but I’m not seeing X in the answer” — check the Trace. If grounding dropped X, the graph probably has a contradicting fact. Fix it in the graph or lower the action to demote.