Hallucination Detection

Detection method

VeldrixAI's hallucination pillar uses ensemble confidence estimation across multiple specialized models. It identifies:

• Invented statistics and numeric claims
• False citations and nonexistent sources
• Factual claims inconsistent with the provided context
• Date and entity inconsistencies

RAG grounding

For maximum accuracy, pass your retrieval context in the context field of the evaluate request. The pillar will verify that response claims are grounded in your retrieved documents.

Limitations

Without a ground-truth context, the pillar estimates plausibility from world-knowledge priors. This is effective for obvious fabrications but may miss subtle domain-specific errors.