RAG for Complex Situational Questions

Vasily Sizov23 May 2026

Problem

Credit specialists at VTB wanted an assistant for 500+ internal documents (PDF, DOCX, XLSX). Their questions are not lookups: a specialist describes a client situation and expects correct conclusions with correct citations. The answer is not stated explicitly and is spread across chapters and documents. Classic RAG did not work, and experts count an answer as wrong even for a missing comma, since errors are expensive in a bank.

Example from the article: "The client is a sole proprietor on the simplified tax system (USN), opened a settlement account 2 months ago, wants a loan for working capital secured by goods in circulation. What requirements apply to them, which documents are needed, and why will this kind of collateral carry a higher interest rate?"

Solution in brief

Classic retrieval stays only as a first pass that finds candidate documents. A Qwen3-30B-class model then filters their chunks (naive LLM RAG). The second version is a Thinking-model agent that searches, checks which parts of the question are still uncovered, and searches again (agentic RAG). It is faster and finds more correct documents and chunks.

Questions this case answers
  • How to answer a multi-part situational question correctly when the answer is assembled from many documents and chunk relevance cannot be judged by vector closeness
  • How to make an agentic search cover every part of a multi-part question and report which parts remain uncovered
  • How to answer complex situational questions when the answer is spread across many regulatory documents
  • How to build a RAG pipeline when classic top-K retrieval with reranking gives poor recall

Tools and methods

  • Docling, deepvk/USER-bge-m3 — chunking and embedder
  • BGE-M3 dense + BM25, merged with RRF — hybrid retrieval in Qdrant
  • Qwen3-30B-class Instruct and Thinking models — chunk filtering, answers, researcher agent
Continue reading — it's free

Google or a one-time email code. No password.