grounding
Making a model answer only from supplied sources, cite them, and refuse when they do not contain the answer.
In more detail
A grounded answer has three duties: every claim comes from a source, each claim cites the passage behind it, and a question the sources cannot answer gets an honest refusal or a handover. Check all three in code, since a model can cite a passage that does not say what it claims.
Related terms
- citation — A pointer from a claim in a model's answer to the source passage that supports it.
- faithfulness — An eval metric for RAG: the share of claims in an answer that the retrieved context supports.
- refusal — A reply in which a model declines to do what was asked, for safety reasons or because it lacks the information.
- hallucination — Output from a model that is fluent and confident but false or unsupported, such as an invented fact, citation, function or package.
Taught in
- How RAG Works And When To Use It — AI Engineering: LLM Apps, RAG & Agents
- Grounded Answers With Citations — AI Engineering: LLM Apps, RAG & Agents
- Guardrails, Privacy And Governance — AI Engineering: LLM Apps, RAG & Agents
- Evaluating RAG And Agents — AI Engineering: LLM Apps, RAG & Agents
References
- Anthropic: Citations
- Lewis et al.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020)