RAG citations face traceability gap

The real issue is architectural: citations need to be treated as first-class data, not decorative links added after generation.

• –Chunk-level IDs, byte or page offsets, document lineage, and transformation history should travel with every retrieved span.
• –Hybrid search can help traceability when BM25 terms provide deterministic anchors that dense retrieval alone may blur.
• –Structured data RAG needs different attribution paths than unstructured text, with row IDs, schema fields, query logs, and source snapshots preserved.
• –Citation-first generation, evidence selection, abstention, and post-generation verification are stronger than asking the model to cite after it has already composed an answer.