RAG Claim Verification

Using AXIMUR to verify claims in retrieval-augmented generation workflows.

Retrieval is Not Verification

RAG systems retrieve relevant passages but do not verify whether retrieved evidence actually supports the AI's claims. Retrieved passages may be irrelevant, outdated, or even contradict the claim entirely.

AXIMUR closes this gap by evaluating the claim against the retrieved evidence itself.

The Verification Pattern

1. Retrieve

Your RAG system retrieves relevant passages from your vector store or knowledge base based on the user query.

2. Generate

The LLM generates a response or claim based on the retrieved context.

3. Verify

AXIMUR verifies the LLM's claim against the retrieved passages:

{
  "claim": "2023 was the hottest year on record globally.",
  "grounding_source": "[Combined retrieved passages from step 1]"
}

4. Act

Based on the verdict, your system can:

• SUPPORTED: Proceed with the claim
• UNSUPPORTED: Flag as contradicted, request revision
• UNCERTAIN: Flag for human review, retrieve more evidence

Use Cases

• Fact-checking AI-generated responses before user display
• Validating agent outputs before irreversible actions
• Citation verification in generated content
• Quality filtering in RAG pipelines

Limitations

AXIMUR only evaluates claims against the grounding source you provide. If your retrieval returns incomplete, outdated, or biased information, the verdict reflects that source — not ground truth. See Limitations for details.

Learn More

• Quickstart — Get started with cURL, TypeScript, Python
• Evidence-Bounded Verification — Understand the verification model
• Pre-Action Verification — Verify before agent actions
• Examples — Full integration examples