RAG Citation Support Verification
Check whether cited passages support each answer claim before a RAG response is shown to users.
An editorial verification workflow that scores claim-evidence support and routes unsupported or ambiguous claims to revision or human review.
- Provider
- Type Safe Workflow
- Reviewed
- September 23, 2026
- Project status
- Community
- Verification
- Needs review
- Categories
- research, development
- Compatibility
- JEV, Laya, Bespoke Nimble, RAG
Inputs
- ✓Atomic claim
- ✓Cited passage
- ✓Document metadata
Outputs
- ✓Support label
- ✓Contradiction label
- ✓Confidence
- ✓Route
Workflow steps
- 1Split generated text into claims that can be judged independently.
- 2Resolve each citation to the exact passage shown to the model.
- 3Judge supported, contradicted, insufficient, or not-applicable for each pair.
- 4Reject missing citations before model evaluation.
- 5Route unsupported, contradicted, and low-confidence claims to revision or review.
- 6Publish only after every material claim passes the policy.
Decision tasks
- ✓Verify
Workflow implementation guide
This editorial pattern requires local evaluation before operational use. Empty tested-version fields mean the directory has not run it against a pinned model release.
Business problem
A retrieved document can be relevant to a topic without supporting the exact claim made in the generated answer.
Who this is for
Teams operating retrieval-augmented assistants, answer engines, or internal knowledge tools.
Model selection notes
Evaluate on domain-specific claim and evidence pairs. The best choice depends on language, evidence length, deployment policy, and error cost.
Tested model versions
Unknown
- Industry
- Knowledge Management, AI Products
- Difficulty
- advanced
- Estimated setup time
- 1 to 2 days including labelled evaluation
- Compatible model candidates
- jev, laya, bespoke-nimble
- Required integrations
- RAG pipeline, Citation mapper
- Recommended model
- No single model recommended before local evaluation
Detailed execution steps
- 1Split generated text into claims that can be judged independently.
- 2Resolve each citation to the exact passage shown to the model.
- 3Judge supported, contradicted, insufficient, or not-applicable for each pair.
- 4Reject missing citations before model evaluation.
- 5Route unsupported, contradicted, and low-confidence claims to revision or review.
- 6Publish only after every material claim passes the policy.
Confidence policy
Use claim-level thresholds calibrated on domain evidence. Contradiction is always blocking; missing or insufficient evidence cannot be rescued by high confidence.
Human review policy
A reviewer handles material unsupported claims, policy-sensitive answers, and disagreements between deterministic citation checks and model judgements.
Privacy notes
Evidence can contain confidential documents. Minimize passages, enforce access controls before retrieval, and avoid logging full inputs.
Cost notes
Cost scales with claim count and evidence length. Deduplicate passages and skip non-material claims only under a documented policy.
Common failure modes
- ✓Long passages contain one matching phrase but do not support the full claim.
- ✓The generated answer combines several facts into one untestable claim.
- ✓Citation access rules are applied after content has already been exposed.
Troubleshooting
- ✓Shorten evidence to the smallest passage that preserves context.
- ✓Split compound claims and retain the mapping to their original sentence.
Structured request example
{"state":{"claim":"The policy covers contractors.","evidence":"This policy applies to employees only."},"questions":{"support":{"type":"choice","criteria":{"supported":null,"contradicted":null,"insufficient":null}}}}Require every material claim to pass
typescript · TypeScript 5.9
type ClaimCheck = {
material: boolean;
label: 'supported' | 'contradicted' | 'insufficient';
confidence: number;
};
export function answerRoute(checks: ClaimCheck[]) {
const unsafe = checks.some((check) =>
check.material && (check.label !== 'supported' || check.confidence < 0.9)
);
return unsafe ? 'revise-or-review' : 'publish';
}Implementation guidance
Keep the declared inputs and outputs visible in code review. Validate every model response before it reaches a database, API, or customer-facing action.