GraphRAG Setup and Usage Guide

Pre-requirements:

Python latest version is giving some errors in installing Gensim and another library, so I recommend you to consider python 3.10 version.

Steps:

1. Open VSCode > Terminal (command prompt) and continue the below commands:

   # Create a new directory
   mkdir Graphrag
   
   # Create a new conda environment with Python 3.10
   conda create -p ./graphragvenv python=3.10
   
   # Activate the created conda environment
   conda activate ./graphragvenv
   
   # Install and update pip
   python -m pip install --upgrade pip
   
   # Install and upgrade the setuptools package
   python -m pip install --upgrade setuptools

2. Install GraphRAG:

   # Install GraphRAG
   pip install graphrag
   
   # If installing GraphRAG raises any error, run the below command
   python -m pip install --no-cache-dir --force-reinstall gensim
   
   # Then install GraphRAG again
   pip install graphrag

3. Initialize GraphRAG:

   python -m graphrag.index --init --root .

   Running the above command, GraphRAG is initiated and creates folders and files.

4. Creating the Deployments of LLM Model and Embedding Model in Azure OpenAI:

   • For embedding, the model is text-embedding-small
   • For LLM model, recommended is GPT-4o.

   I have tried with GPT-35-turbo-instruct which is not working, raising some issues on creating the graphs at the end. I recommend you to use GPT-4o, and the model is very good at providing accurate answers. You can also check which other models

   After creating the deployments, you need to make some changes in the settings.yaml file.

5. Configure OPENAI_API_KEY:

   .env is also created when you initialize GraphRAG, you can configure your OPENAI_API_KEY there.

6. Make Changes in settings.yaml File:

   Go to the llm section:

   • Change the api_key from ${GRAPHRAG_API_KEY} to ${OPENAI_API_KEY}
   • Change type from openai_chat to azure_openai_chat
   • Change model to the model name that you deployed in Azure OpenAI.
     • In my case, I have taken GPT-4o, you can try with other models as well.
   • Uncomment the api_base and replace that URL with the endpoint of the Azure OpenAI
     • In my case <your_azure_openai_resource_group_name> is used, you can give any name to the instance.
   • Uncomment the api_version, no need to make changes in that, you can use the default api_version
   • Uncomment the deployement_name, and replace that with the deployment you have given to the model while creating the deployment.
     • In my case, I have taken <deployment_name_of_model>

   In the same settings.yaml file, go to the embeddings section and make some changes:

   • Change the api_key from ${GRAPHRAG_API_KEY} to ${OPENAI_API_KEY}
   • Change type from openai_embeddings to azure_openai_embeddings
   • Change model to the model name that you deployed in Azure OpenAI.
     • In my case, I have taken text-embedding-small, you can try with other models as well.
   • Uncomment the api_base and replace that URL with the endpoint of the Azure OpenAI
     • In my case <your_azure_openai_resource_group_name> is used, you can give any name to the instance.
   • api_version, you can comment that or you can uncomment and use that. It won't raise an error
   • Uncomment the deployement_name, and replace that with the deployment you have given to the model while creating the deployment.
     • In my case, I have taken <deployment_name_of_model>

   That's it. Now save the changes that you made in the settings.yaml file.

7. Add Input File:

   You need to add an input text file, to do that first create a folder by running the below command:

   mkdir input

   Now in this input folder, you need to add the .txt file (input text data file).

8. Run GraphRAG to Create Graphs on the Data:

   python -m graphrag.index --root .

   This command will run the GraphRAG and create the parquet files. This will convert all your text data into entities and relationships graphs. You can check that in output > last folder > artifacts folder (you have parquet files, which are converted data into graphs).

9. Evaluate Our RAG Model:

   Now we need to test our RAG model, to do that we need to ask questions. We are doing things in the command prompt, we need to ask questions to our model along with a command. To ask a question, you need to write a command and then the question:

   python -m graphrag.query --root . --method local/global "Your_Question"

   When you run this command, the model uses either local (if written local) or global (if written global) to answer the question.

   To confirm the model is accurate, we can copy any word from the output of the model and find it in the text file:

   • Copy the word you need to check
   • Go to the text file (input file) and hit ctrl+f, a search bar will open
   • Paste that copied word in this search bar, use the arrows to navigate to the words in the text file.