Flexible GraphRAG

Flexible GraphRAG is a self-hosted platform for teams building document search and AI chat that use both text and relationships between entities. It extracts knowledge graphs from documents and combines graph retrieval with vector similarity and full-text search. You can run it standalone or in Docker. It's open source under Apache 2.0.

Model and parser choices determine which processing happens locally. Ollama and vLLM support local LLM use, while Docling and LiteParse parse documents locally. LlamaParse sends documents to a cloud API; OpenAI, Anthropic Claude and other hosted model providers are optional alternatives. Docling supports CPU processing and acceleration on CUDA or Apple Silicon GPUs, plus OCR for scans.

The platform accepts uploaded files and connects to repositories such as Alfresco, Nuxeo and SharePoint, along with cloud storage and web content. Automatic change sync keeps supported sources reflected in the search and graph databases. Parsers cover documents, images and audio, with format support varying by parser.

Database choice is a central part of the design. It works with Neo4j and other property graph stores, RDF stores such as Ontotext GraphDB, and vector databases including Qdrant and PostgreSQL pgvector. Schemas and OWL/RDFS ontologies can guide entity and relationship extraction.

Browser interfaces are available in Angular, React and Vue. REST APIs and an MCP server expose ingestion, search and AI queries to other applications, including Claude Desktop. Optional Langflow flows provide visual pipeline editing; CocoIndex offers an alternative ingestion engine with memoization and delete reconciliation.