02.02 · Lecture · Free
Grounding: An AI Answer Is a Search Result With a Writer on Top
Trace a generative answer back through grounding to the retrieved documents it was built from, and say which stage a missing citation actually failed at.
Curated for this lesson1/3
How AI Answers Are Built
Stanford CS25: V3 I Retrieval Augmented Language Models
Stanford lecture specifically on retrieval-augmented language models, matching the search-plus-writer mental model.
A grounded AI answer is best read as retrieval plus generation: documents are found, selected, passed into context, then rewritten into a response. Unsupported claims can fail at indexing, retrieval, ranking, prompt assembly, generation, or citation rendering, so debugging starts by tracing the final text back to the evidence pipeline.
What this lesson answers
- how does grounding work in AI answers
- why is an AI citation missing
- how to debug retrieval augmented generation output
Notes
A grounded AI answer is not just something the model “knows.” Think of it as two systems connected together: a search system finds relevant documents, then a language model writes an answer using the retrieved text as evidence. The answer may sound fluent either way, but grounding is what gives you a trail from the final sentence back to a source document.
The practical pipeline is: turn the user question into a search query or embedding, retrieve candidate documents, rank or filter them, pass selected chunks into the model’s context, and ask the model to answer with citations. When debugging, follow that chain backwards. Did the right document exist in the index? Was it retrieved? Was the relevant chunk included in the prompt? Did the model use it? Did the citation formatter attach the right source?
A common misconception is that a missing citation means the model hallucinated. That can be true, but it is only one possible failure. The retrieval stage may have failed to find the document, the chunking stage may have cut away the useful passage, the reranker may have dropped it, the prompt may have omitted it, or the generator may have answered from general language-model behavior instead of the provided context.
After this lesson, you should be able to inspect an AI answer like a production trace. For any unsupported claim, ask which stage lost the evidence: indexing, retrieval, ranking, context assembly, generation, or citation rendering. That habit turns “the AI was wrong” into a debuggable system failure.
Common questions
- What does grounding mean for an AI answer?
- Grounding means the answer is tied to retrieved source material rather than produced only from the model’s internal patterns. A search component finds candidate documents, selected passages are placed into the model context, and the generated response should be supported by that evidence. The important property is traceability from claim to source.
- Does a missing citation always mean hallucination?
- No. A missing citation can come from several stages before generation. The source may not have been indexed, the right passage may not have been retrieved, ranking may have discarded it, context assembly may have left it out, or the citation layer may have failed even though the model used valid evidence.
- How should I debug an unsupported AI claim?
- Work backwards from the claim. Check whether the source exists in the index, whether retrieval returned it, whether the useful passage survived chunking and ranking, whether it reached the model prompt, whether the model actually used it, and whether the citation formatter pointed to the correct document.