Attribution links a generated claim to an identified source passage that independently supports it. A response without supporting evidence is the claim to distrust; a cited passage lets a reviewer verify one sentence instead of reconstructing the answer. section 1.3.5
Retrieval makes attribution possible but does not guarantee it. Retrieval finds context, while generation uses that context to produce a grounded response; attribution asks whether each resulting statement is actually supported. section 1.3.5 section 1.2
Useful attribution checks operate at several stages:
- Provide source passages alongside answers so reviewers can check the evidence. section 1.3.5
- Grade retrieved documents before generation to detect weak evidence early. section 1.3.5
- Split an answer into claims and verify each claim against the supplied sources. section 1.3.5
Accuracy is not attribution
Across 4,326 searches, an AI search system’s factual accuracy rose from 85% to 91%, while unsupported citations among correct answers increased from 37% to 56%. A separate audit found only 51.5% of generated sentences fully supported by citations. section 1.3.5
Figure 1.1 The RAG workflow: from user queries to grounded responses through retrieval and generation.

Trace how query processing, knowledge-base search, retrieved context, and generation combine to produce a response that can carry source attribution. section 1.2


