Learning outcomes
- Classify factual failures
- Measure evidence support
- Select grounding controls
Mental model
A hallucination is an unsupported or incorrect model claim. Grounding reduces the allowed claim space by connecting outputs to trusted evidence, but retrieval alone does not guarantee support.
Theory
Failures can originate in missing retrieval, irrelevant retrieval, conflicting sources, synthesis beyond evidence, stale data, or incorrect tool results. Evaluate these stages separately. Useful controls include authoritative source selection, abstention, claim decomposition, citation support checks, deterministic tools, and human escalation.
Alternatives and trade-offs
Closed-book generation is simplest, RAG supplies documents, tools supply structured truth, and extractive responses reduce transformation when precision dominates fluency.
Failure modes and misconceptions
Do not use a citation count as factuality, force an answer when evidence is absent, or label every disagreement a model failure without checking source quality.
Knowledge check
Which measurements distinguish a retrieval failure from a generation failure?
Decision scenario
When a benefits answer is unsupported, record whether the correct passage was retrieved before changing the prompt or model.
Relationships
Grounded Generation and Citations
Grounding failures are evaluated against retrieved evidence and claims.
prerequisiteLLM Evaluation
Grounding quality needs explicit cases metrics and rubrics.
Primary sources
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks - arXiv, verified 2026-07-16
- Evaluation Best Practices - OpenAI, verified 2026-07-16