# DocMind AI

A local-first Streamlit application for document analysis that combines hybrid retrieval, optional knowledge-graph search and multi-agent coordination over local large language models.

DocMind AI is a local-first document analysis application built on LlamaIndex and LangGraph, running local LLMs through Ollama, LM Studio, llama.cpp or vLLM. It parses PDFs and common office and HTML formats with Docling, performs OCR on scanned pages, and blocks remote language model endpoints by default so an explicit choice is needed to cross the local trust boundary.

Retrieval combines dense and sparse vector search in Qdrant with server-side fusion, keyword search, optional knowledge-graph retrieval, and multimodal page retrieval, all selected by a router query engine; a cross-encoder reranks text results and a separate model reranks visual ones. A LangGraph supervisor coordinates four agents, a planner, retrieval, synthesis and validation agent, to answer document questions, and page images are rendered and stored as content-addressed artifacts with optional encryption.

Ingestion is deterministic and cached, and the app can snapshot its indexes with corpus and configuration hashes for reproducibility. It is written in Python and licensed under MIT.

Website: https://github.com/BjornMelin/docmind-ai-llm