GraphRAG — Getting Started

Before you read this: GraphRAG assumes you already understand ordinary RAG — embeddings, a vector database, chunking, a chat model answering from retrieved passages. That is the whole content of RAG Topics. This section only covers what is different about GraphRAG.

A GraphRAG Topic is the graph-based sibling of a RAG Topic. Where a RAG Topic only stores text pieces (chunks) in a vector store, a GraphRAG Topic also extracts the entities and relationships those chunks describe into a graph database. It groups that graph into topical communities and writes a short summary of each one. For every chat question, a router model then chooses how to answer: a precise entity lookup (LOCAL), a broad overview (GLOBAL), both combined (HYBRID), a search for how two things are connected (PATH), or no search at all (DIRECT or ATTACHMENT).

It is a separate resource type from a RAG Topic. It has its own list, wizard, detail page, chat and permissions (GRAPHRAG_* instead of TOPIC_*). Creating one does not touch your existing RAG Topics, and the two can exist side by side.

Good to know. Uploading is slower and uses more model calls than a RAG Topic, because a model reads every text piece to find entities. Try a GraphRAG Topic with a small set of your own documents first. Check that its answers fit your questions before you load everything.

What you need before you start

Everything a RAG Topic needs, plus a graph database:

Needed Where it comes from Notes

An AI server with a chat model

Admin → AI Servers

Writes the answers. Optional model roles can reuse it

An AI server with an embedding model

Same

Text pieces, entities and community summaries are all embedded with it

A registered graph database

Admin → Graph Databases

Neo4j, Memgraph, a generic Cypher server or CYROCK.DB. Required: there is no built-in graph store, and the wizard does not let you continue without one

A registered vector database

Admin → Vector Databases

Required, exactly as for RAG Topics: there is no built-in vector store. You pick it in step 1 of the wizard

A Docling server, if you will upload PDF, Word or similar files

Admin → Docling Servers

Needed once binary documents are an allowed file type

A vision model, if pictures should be described

Any AI server

Needed for picture handling (see File types, Docling and pictures). Every provider can serve this role

A free host port for the Topic

Any unused port on the host

Required. It is the only way in to the Topic

The graph database is the one new requirement. Register it under Admin → Graph Databases. The dialog works like the one for vector databases: host, port, credentials and a connection test. See Connect a Graph Database.

The path from zero to a working GraphRAG Topic

  1. Read Why GraphRAG — five minutes to understand what a graph adds to plain RAG, and whether your content benefits from it.

  2. Register a graph database under Admin → Graph Databases. Neo4j is the easiest to try locally: one docker run gives you one. Register a vector database under Admin → Vector Databases too. A GraphRAG Topic needs both.

  3. Run the four-step wizard. Creation step by step walks through it field by field.

  4. Start the Topic and upload documents on its Data Upload tab. Uploading is slower than for plain RAG, because a model reads each text piece to find entities. See the ingestion workflow for what happens underneath.

  5. Wait for communities. A GraphRAG Topic is not fully ready the moment an upload finishes. Community summaries are what make overview questions answerable. The Topic rebuilds them on the refresh cycle you set in the wizard. You can also start a rebuild yourself on the Actions tab, in the section Communities. See Configuration and functions.

  6. Chat. Ask a narrow question about one thing, and then a broad question about all documents. Notice that they are answered differently. That difference is the whole point of the router and its search modes. See Retrieval workflows and the GraphRAG Chat Window.

Related