Enterprise RAG systems often encounter a challenge where answers point to other sections (e.g., 'see Section 7.2') rather than providing the direct information. This article introduces 'loop engineering' to systematically follow these internal pointers.

The process involves a two-pass system. The first pass parses the document and performs retrieval. If the generation step identifies unresolved references, an orchestrator signals for a second pass. This second pass resolves the references using parsed relational tables, re-retrieves the linked content, and then re-runs generation with the complete context.

References can occur during parsing, retrieval, or generation. The proposed fix uses feedback fields in the generation output to flag unresolved references. The orchestrator then utilizes this signal to resolve the reference and re-fetch the necessary information. The system is demonstrated using the 'Attention Is All You Need' paper, where an initial retrieval might miss crucial details found in a referenced table.

The parsing stage cheaply identifies native PDF links and section anchors. Other references, like 'see Table 3 row (E)', are detected on demand during the second pass using regex or LLM analysis. This ensures efficiency by only performing costly analysis when necessary, thereby improving the completeness and accuracy of RAG system responses.