GraphRAG is a Python project that uses graph-based methods for information retrieval. It uses language models and embeddings to create and interact with a graph of data.
The project is structured as follows:
main.py: The main entry point of the application.Graph_Generation/: Contains the code for generating the graph.Graph_Retrieval/: Contains the code for retrieving information from the graph.config.yaml: Configuration file for the application.requirements.txt: Lists the Python dependencies required for the project.
- Clone the repository
git clone https://github.com/Sathvik21S21Rao/KnowledgeGraph.git-
Update the
config.yamlfile, with relevant API keys and models to be used. -
Create graph and model directory in the working directory
mkdir graph
mkdir model
mkdir node_data- Build the docker image in the working directory
docker build -t graphrag .- Run the docker container in interactive mode
docker run -it --mount source=VOLUME_NAME, target=/app graphrag- In case you are running an Ollama model locally
docker run -it --network host graphrag- In case you want to run it in your system
pip install -r requirements.txtThe application's behavior can be configured by modifying config.yaml. This includes the paths to the graph, node data, and community data, as well as the settings for the language model and embeddings.
The application supports different language models and embeddings, which can be configured in config.yaml. The supported language models are OpenAI, Ollama, and Groq. For OpenAI and Groq fill in the api_key field.
Graph Generation is done with the help of the LLM, by identifying entities and relationships. The prompts for generating graphs have been taken from Microsoft's graphrag with modifications. The graphs are then analyzed using node2vec, which produces graph embeddings. The graph, the corresponding node names, and the embedding models are stored. Girvan Newman technique has been used to identify first-level communities, which are summarized.
Graph Retrieval is used during query. The LLM is used twice: once for extracting relevant nodes from the graph and second for answering the query. After extracting relevant nodes, the node2vec model is used to enrich the context by retrieving nodes that are similar to the present nodes (using graph embeddings). If the nodes belong to more than 40% of the communities, then the context is switched to the community summaries. Otherwise, the context format is as follows: Nodes and their Description | Edges and their Description.
This is done to ensure the size of the context is optimized and relevant information stays intact.
Everytime the interactive session is run, the program actively checks for additions into the data file. In case additions are found, the graph will be updated accordingly. The communities will be updated in case of:
- New nodes in the community
- Updation of existing nodes in the community
- A creation of an entirely new community
The updation community prompt has been optimised so as to return minimum number of output tokens.(The entire community summary is not regenerated)
When running main.py, you will be asked whether you want to create a new graph. If you choose not to create a new graph, the program will attempt to load a graph from the specified path.
Once the application is running, you can interact with it by typing queries into the console. The application will respond with information retrieved from the graph based on your query. To end the conversation, type 'exit'.
- An interactive UI for seamless user experience
- Integrating vector search along with graph search
- Allowing users to maintain multiple graphs.
- Shifting from in memory graph computation to disk based retrieval
- Built using the data in testing_data.py
- The nodes of the same colour are part of the same community
