Installation
To use Agentic GraphRAG, install agentic-graphrag from your Python package manager, for example pip:
pip install agentic-graphrag
If you are using uv, create a virtual environment and install the package into it:
uv venv
uv pip install agentic-graphrag
To install uv, see the installation docs.
Agentic GraphRAG needs Python 3.11 or newer. It also needs a Neo4j database, version 5.23 or newer, to store the graph. The Quickstart shows how to get one for free.
The base install includes the loaders, the chunkers, the data models, and the default embedder. Storage, extraction, and the agent are extras.
Alternative PyTorch distributions
The embed-local, extract, docling, and chunk-neural extras depend on PyTorch. Depending on your hardware, pick a different distribution of torch, for example a CPU-only build or a build for another accelerator. The PyTorch install page lists all the ways to install torch.
When you install with uv, add --torch-backend=auto. uv looks for a GPU driver. It installs the build that matches the driver, or the CPU build when it finds no GPU.
uv pip install "agentic-graphrag[neo4j,extract]" --torch-backend=auto
To always install the CPU build, use --torch-backend=cpu. This flag is a uv preview feature. It needs uv 0.6.9 or newer.
What the base install includes
The base install has the parts of Agentic GraphRAG that need no database, no GPU, and no LLM client:
- The data models, such as the graph schema and the search result.
- Loaders for files, directories, and raw text, including Markdown, HTML, and CSV.
- The built-in chunkers: recursive, sentence, token, heading, parent-child, and turn-window.
- The default embedder, which runs ONNX embedding models on the CPU.
- The OpenTelemetry API calls that trace each step. Without the
observabilityextra, these calls do nothing.
To build and search a graph, you also need a graph store and an extractor. Each one is an extra.
Available extras
Other parts of Agentic GraphRAG need more third-party packages. These packages install only if you select them as extras. If an extra is missing, Agentic GraphRAG raises an error that names the extra to install.
The following table lists the extras. Install them with pip install "agentic-graphrag[NAME1,NAME2]" or uv pip install "agentic-graphrag[NAME1,NAME2]".
| Extra | Description |
|---|---|
neo4j | Installs the Neo4j graph store and its native vector index. |
embed-local | Installs Sentence Transformers, to embed with models that the default embedder does not support. |
extract | Installs GLiNER 2.5, to run entity extraction on your machine. |
llm | Installs BAML, to extract entities and review matches with an LLM. |
agents | Installs the LangGraph runtime for the question-answering agent. |
docling | Installs Docling, to parse PDF, Word, PowerPoint, image, and XML files and to chunk by layout. |
qdrant | Installs the Qdrant vector store and BM25 sparse embeddings. |
weaviate | Installs the Weaviate vector store. |
milvus | Installs the Milvus and Zilliz vector store. |
observability | Installs the OpenTelemetry SDK and OTLP exporter. |
community | Installs Leiden community detection. |
eval | Installs the DeepEval and AgentEvals metrics in agrag.eval. |
chunk-semantic | Installs the semantic chunker. |
chunk-neural | Installs the neural chunker. |
chunk-code | Installs the code chunker. |
Extras by goal
| Goal | Extras |
|---|---|
| Build a first graph (the Quickstart) | neo4j,extract |
| Also ask questions with the agent | neo4j,extract,agents |
| Read PDF and Office files | Add docling |
| Extract with an LLM instead of a local model | neo4j,llm |
| Search with a dedicated vector store | Add qdrant, weaviate, or milvus |
| Find communities of related entities | Add community |
| Record traces | Add observability |
| Measure answer quality | Add eval |
Next steps
Follow the Quickstart to build your first graph.