Grounding
Grounding ties generation to external evidence. In RAG, that evidence is retrieved text; in agents it may be tool output or environment state. Citations are the visible surface of grounding, while hallucination mitigation tests unsupported claims. A grounded answer should be constrained by what the system actually observed, not by what the model may remember from training.
The grounded-answer contract
A grounded answer contract has three parts: source selection, claim generation, and support checking. Each factual claim should be derivable from a source span or tool observation included by context construction. If no source supports the answer, the model should say so rather than fill gaps from parametric memory.
| Layer | Question | Example failure |
|---|---|---|
| Retrieval | did the system fetch the right evidence? | correct policy exists but was not retrieved |
| Context construction | was the evidence visible to the model? | relevant span was truncated or buried |
| Generation | did the answer stay within the evidence? | model adds a condition not present in sources |
| Verification | can each claim be traced to a span? | citation points to a related but non-supporting passage |
Grounding is stronger than adding citations after the fact. The citation should be a support relationship between a claim and evidence, not a decorative link to a broadly relevant document.
Support relationships
| Relationship | Meaning | Example |
|---|---|---|
| Entailed | the span directly supports the claim. | claim: “approval required above 500 EUR”; span states the same threshold. |
| Contradicted | the span says the opposite or a different value. | claim says two days; span says five business days. |
| Related | same topic but not enough evidence. | span explains refunds but not approval thresholds. |
| Missing | no retrieved span supports the claim. | answer mentions a policy date absent from context. |
Grounding validators should distinguish these cases. Related evidence is the most dangerous because it makes the answer look cited while not actually supporting the claim.
A claim-support record
{
"claim": "Manager approval is required above 500 EUR.",
"source_id": "policy-7",
"span": "Manager approval is required above 500 EUR.",
"support": "entailed"
}This source-level artifact gives rag evaluation something concrete to score.
Grounded Refund Answer
User question:
Does this 700 EUR enterprise refund require manager approval?Retrieved source:
For enterprise accounts, manager approval is required for refunds above 500 EUR.
Finance approval is required above 5000 EUR.Grounded answer:
Yes. A 700 EUR enterprise refund requires manager approval because it is above the 500 EUR threshold. It does not meet the 5000 EUR finance-approval threshold. [policy-7]This answer is grounded because both the approval requirement and the non-requirement for finance approval are derived from the retrieved span. A less grounded answer would say “approval usually takes two days” if the span does not mention timing.
Evaluation
Grounding is evaluated at the claim level. Split an answer into factual claims, attach source spans, and score whether each claim is entailed, contradicted, related, or missing. Aggregate metrics should report unsupported-claim rate and citation-support rate separately. A system can have many citations and still be poorly grounded.
Caveats
Grounding can fail when retrieval misses the right passage, when sources conflict, or when the answer overgeneralizes beyond the span. It can also fail silently when the retrieved text is stale or when a source is authoritative for one jurisdiction, product version, or time period but not another. Strong grounding therefore needs metadata such as date, version, tenant, and source authority, not only text snippets.
References
Nav