This repo contains recipes and notebooks to experiment with Graph RAG strategies using Kùzu. Feel free to clone the repo and adapt the code to your own projects!
Graph RAG can be thought of as a suite of methodologies that incorporate a knowedge graph (or more simply, a graph) into the retrieval process to enhance the relevance and factual accuracy of the generated responses.
In building a Graph RAG application, we must first ask the following primary questions:
- What is the graph? What do the nodes and edges represent?
- How is the retrieval process different from traditional vector-based retrieval in RAG?
There are multiple strategies that can be used to enhance the retrieval context using a graph. Some of them are covered in the examples in this repo.
Like in any other RAG, there are two key stages in building a Graph RAG application:
- Indexing: This stage involves extracting explicit entities and their relationships from the data and building a graph, and inserting vector embeddings of the chunks of text to build a vector index.
- Serving: This stage involves sending the user query to the graph and/or the vector index, to enhance the retrieval quality and answer relevance.
For Graph RAG, we will largely be using the Python API and testing out ideas using Jupyter notebooks
or Python scripts. You can manage dependencies using requirements.txt files, installed via pip,
or the uv package manager (recommended).
The example below shows how to set up a virtual environment using uv.
# Set up a virtual environment using the uv package manager
# https://github.com/astral-sh/uv
uv venv
source .venv/bin/activate
# Install kuzu
uv pip install -r requirements.txtYou can then create a new ipykernel, so that you can use the virtual environment in the notebook.
ipython kernel install --user --name=rag_envLocate the new kernel in the list of kernels in Jupyter and select it to begin installing packages and running the notebooks.