Faithfulness
also called Grounding
Sticking to the material it was given.
The degree to which an answer is supported by the supplied context rather than invented or imported from the model’s own priors. The core quality metric for RAG systems.
Unfaithful answers are especially dangerous because the source was right there. The fix is usually context hygiene: remove contradictory or stale passages and instruct the model to say when the answer is not present.
Faithfulness is measured against supplied material, not against the world, and that distinction makes it the practical quality metric for retrieval systems: you can grade it automatically by checking whether claims are supported by the retrieved text, even when you cannot verify them against reality.
Correct but unfaithful
A model answers from its training data rather than the document it was given.
The answer may be true and still be a bug — the document said something different, and you are now shipping a claim you cannot trace. Faithfulness failures hide behind correct-looking output, which is why they need their own check.
- paperRetrieval-augmented generation for knowledge-intensive NLP tasks ↗
The original RAG formulation: combine a parametric model with a non-parametric dense index, so answers can be grounded in retrieved text.
- docsGrounding overview ↗
Grounding responses against search results or your own corpora, returning support citations that tie individual claims back to sources — the API shape of provenance.
- standardLLM09: Misinformation ↗
Treats overreliance as the risk: the failure is users accepting confident output without verification, so the mitigation is validation and visible limitations, not just better models.
Hallucination
Fluent output that is not true.
RAG · Retrieval-Augmented Generation
Look things up, then put what you found in the prompt.
Grounding
Tying outputs to evidence the user can check.
Evaluation · Evals
How you know the system is actually working.
Context Rot
A long session slowly getting worse.