attribution

Attribution traces which retrieved sources support each part of a generated answer, enabling verification. **Motivation**: Users need to verify AI claims, trust requires transparency, citations enable fact-checking. **Implementation approaches**: **Post-hoc**: Generate answer, then match statements to sources via NLI/similarity. **Inline generation**: Train/prompt model to cite sources as it generates [1], [2] style. **Structured output**: Model outputs (statement, source_ids) pairs. **Citation quality**: Precision (cited sources actually support claim), recall (all claims have citations), verifiability (human can check). **Challenges**: Generated text may paraphrase sources, combining information from multiple sources, hallucinated citations. **Evaluation**: ALCE benchmark, human evaluation of citation quality. **Tools**: LangChain source tracking, LlamaIndex citation engine. **UI considerations**: Display sources alongside text, link to original documents, highlight supporting passages. **Best practices**: Retrieve high-quality sources, verify citations before presenting, allow users to see source context. Attribution builds trust and enables human-AI collaboration for accuracy.