AI and LLM integrations#
Overview#
AgensGraph is a common backing store for retrieval-augmented generation, because a knowledge
graph and its embeddings can live in one database. The graph is tables, and a property is
jsonb, so pgvector indexes an embedding stored on a vertex.
One query can therefore find the semantically closest documents and then walk the graph out from them. GraphRAG needs exactly that.
The integrations live in their own manual: AgensGraph AI. Each has its own installation and usage pages there; this page says what each one is for.
Every one of them runs on the AgensGraph Python driver, which is built on psycopg 3 and runs Cypher through it. So a vertex, an edge or a path reaches the framework as an object rather than as text, whichever framework you pick. The 0.3.1 integrations require version 2.0 or later of the driver. See Drivers and client libraries for the driver on its own.
pgvector handles the embeddings and is not bundled, so you install it into the database
once. See Vector search.
Source for every integration: the agensgraph-ai repository, which also ships runnable demos.
LangChain#
A graph store, a pgvector-backed vector store, chat-message history, a LangGraph checkpointer, and an LLM graph transformer.
LlamaIndex#
PropertyGraphIndex and VectorStoreIndex stores, with hybrid search, metadata filtering and
AgensGraph-tuned Text2Cypher.
LightRAG#
All four LightRAG storage roles on one database, with workspace multi-tenancy.
cognee#
Graph and vector adapters for the cognee AI-memory framework.
MCP servers#
Three Model Context Protocol servers — Cypher, memory and data modeling — that expose AgensGraph to clients such as Claude Desktop.
Without a framework#
If you are building the retrieval yourself rather than through a framework, the three search guides are the place to start. Vector, full-text and hybrid search each end with a GraphRAG section that combines the search with a traversal in one statement.