citation accuracy

**Citation accuracy** is the **measurement of whether cited sources in a generated answer actually support the specific claims they are attached to** - it is a core trust metric for retrieval-augmented systems. **What Is Citation accuracy?** - **Definition**: Degree of correctness in claim-to-source references included in model outputs. - **Evaluation Unit**: Assessed at statement level rather than only at whole-answer level. - **Failure Modes**: Includes wrong source links, mismatched passages, and irrelevant citations. - **System Role**: Connects retrieval evidence to user-visible verification paths. **Why Citation accuracy Matters** - **User Trust**: Accurate citations let users verify claims quickly and confidently. - **Hallucination Detection**: Citation mismatch is a strong signal of unsupported generation. - **Compliance Readiness**: Regulated environments require defensible references for key outputs. - **Product Quality**: High citation fidelity improves perceived reliability of AI assistants. - **Debug Value**: Citation errors reveal retrieval, ranking, or grounding defects. **How It Is Used in Practice** - **Claim Extraction**: Break answers into atomic claims and validate each against cited passages. - **Automated Scoring**: Use verifier models plus human spot checks for citation support labeling. - **Prompt Guardrails**: Force passage IDs and source spans into generation constraints. Citation accuracy is **a key acceptance metric for trustworthy RAG products** - improving citation accuracy directly raises factual transparency and user confidence.