Enterprise RAG pipelines often return incomplete answers when retrieved passages contain cross-references like 'see Section 7.2' instead of the actual data. This post presents a loop engineering pattern that resolves such references systematically: the generation schema is extended with `pending_references` and `answer_completeness` fields; when the LLM flags an unresolved reference, an orchestrator triggers a second pass that resolves the reference deterministically against relational tables (toc_df, object_registry) or via a small LLM call for ambiguous phrasings, then re-retrieves the target region and re-generates a complete answer. The approach deliberately keeps parsing and retrieval cheap by default (lazy reference extraction, top-1 retrieval) and expands only on typed signals from structured output, avoiding confidence-score heuristics. Four reference types are covered: section, table/figure, conditional clause, definitional, and external references.
