🔍 AI Hallucination & Fact Check

Contextual grounding and hallucination risk evaluator for RAG pipelines. Pinpoints unverified entities, phantom dates, numbers, and ungrounded lexical assertions.

Evaluation Scenario:

Ground Truth Source (RAG Context Chunk):

Generated LLM Response Claim:

Faithfulness Score

Grounding Ratio

Hallucination Risk

HIGH RISK

Confabulation metric

Ungrounded Tokens

Novel Named/Numeric Terms

⚠️ Flagged Unsupported Entities & Numbers (Absent in Context):

Self-Correction Mitigation: If hallucination risk is moderate or high, inject an explicit CoT prompt constraint: "Answer strictly using only the facts mentioned in the context. For any fact not explicitly mentioned, state: 'The provided documents do not state this.'"