Dispatcher orchestrates TOC retrieval, listing aggregation, and synthesis patterns in RAG systems. Each pattern has bounded iteration logic to handle complex questions like NIST Cybersecurity Framework queries. The pdf_qa_loop function composes these patterns into a single workflow, using typed objects (Parsed Question, Document Profile) to manage feedback loops.

Unlike agentic approaches, control stays in code for auditability. The dispatcher decides which patterns activate, while loops manage iterations based on feedback fields. This structured approach avoids special cases, scaling to multi-intent requests.

The system maintains modularity, with bricks (parsing, retrieval) unchanged during updates. Real-world questions trigger multiple patterns simultaneously, requiring the dispatcher to coordinate sequences. The pdf_qa_loop runs in production, with companion code on GitHub (doc-intel/notebooks-vol1) for reproducibility.

This workflow represents a five-rung evolution from basic to agentic RAG, stopping two rungs short of full model control.