An agentic RAG pipeline that stops to ask for directions

What it does This repository is a learn-by-building kit for Agentic RAG. It ingests PDFs, converts them to Markdown, and indexes them hierarchically in Qdrant—small child chunks for search precision, large parent chunks for answer context. A LangGraph-orchestrated agent handles queries through a four-stage pipeline: conversation memory, query clarification, parallel multi-agent retrieval, and final response aggregation.

The interesting bit The pipeline doesn’t just retrieve; it interrogates. Ambiguous queries are rewritten or bounced back to the user for clarification before any vector search happens. Complex questions are decomposed and farmed out to parallel agent subgraphs that self-correct if their retrieval looks thin, then compress context to avoid stuffing the window. That orchestration is usually the hard part, and here it is laid out as swappable modules.

Key highlights

• Hierarchical indexing: searches precise child chunks, then fetches richer parent chunks to ground answers
• Multi-agent map-reduce: splits complex queries into parallel sub-queries handled by dedicated agent subgraphs
• Self-correction and context compression: agents re-query when results are insufficient and prune redundant context across loops
• Provider-agnostic LLM backend: swap between Ollama, OpenAI, Anthropic, or Google via a single line of LangChain configuration
• Observability: Langfuse tracing for LLM calls, tool usage, and graph execution
• Human-in-the-loop clarification: detects vague inputs and pauses to ask the user rather than hallucinating an interpretation

Caveats

• Local models under 7B parameters may ignore retrieval instructions or hallucinate, so Ollama users should size up accordingly
• Model identifiers change frequently; the README warns that you will need to verify current names against provider docs before deploying

Verdict Grab this if you are trying to understand how agent orchestration, hierarchical retrieval, and query clarification fit together in a real pipeline. Skip it if you already have a battle-hardened RAG system and just need a drop-in vector search replacement.