Table of Contents

What is KG-RAG

Example use case of KG-RAG

• Prompting GPT without KG-RAG
• Prompting GPT with KG-RAG
• Example notebook for KG-RAG with GPT

How to run KG-RAG

• Step 1: Clone the repo
• Step 2: Create a virtual environment
• Step 3: Install dependencies
• Step 4: Update config.yaml
• Step 5: Run the setup script
• Step 6: Run KG-RAG from your terminal
  • Using GPT
  • Using GPT interactive mode
  • Using Llama
  • Using Llama interactive mode
• Command line arguments for KG-RAG

Citation

What is KG-RAG?

KG-RAG stands for Knowledge Graph-based Retrieval Augmented Generation.

Start by watching the video of KG-RAG

KG_RAG_schematics.mov

It is a task agnostic framework that combines the explicit knowledge of a Knowledge Graph (KG) with the implicit knowledge of a Large Language Model (LLM). Here is the arXiv preprint of the work.

Here, we utilize a massive biomedical KG called SPOKE as the provider for the biomedical context. SPOKE has incorporated over 40 biomedical knowledge repositories from diverse domains, each focusing on biomedical concept like genes, proteins, drugs, compounds, diseases, and their established connections. SPOKE consists of more than 27 million nodes of 21 different types and 53 million edges of 55 types [Ref]

The main feature of KG-RAG is that it extracts "prompt-aware context" from SPOKE KG, which is defined as:

the minimal context sufficient enough to respond to the user prompt.

Hence, this framework empowers a general-purpose LLM by incorporating an optimized domain-specific 'prompt-aware context' from a biomedical KG.

Example use case of KG-RAG

Following snippet shows the news from FDA website about the drug "setmelanotide" approved by FDA for weight management in patients with Bardet-Biedl Syndrome

Ask GPT-4 about the above drug:

WITHOUT KG-RAG

Note: This example was run using KG-RAG v0.3.0. We are prompting GPT from the terminal, NOT from the chatGPT browser. Temperature parameter is set to 0 for all the analysis. Refer this yaml file for parameter setting

bbsyndrome_without_kgrag.mov

WITH KG-RAG

Note: This example was run using KG-RAG v0.3.0. Temperature parameter is set to 0 for all the analysis. Refer this yaml file for parameter setting

bbsyndrome_with_kgrag.mov

You can see that, KG-RAG was able to give the correct information about the FDA approved drug.

How to run KG-RAG

Note: At the moment, KG-RAG is specifically designed for running prompts related to Diseases. We are actively working on improving its versatility.

Step 1: Clone the repo

Clone this repository. All Biomedical data used in the paper are uploaded to this repository, hence you don't have to download that separately.

Step 2: Create a virtual environment

Note: Scripts in this repository were run using python 3.10.9

conda create -n kg_rag python=3.10.9
conda activate kg_rag
cd KG_RAG

Step 3: Install dependencies

pip install -r requirements.txt

Step 4: Update config.yaml

config.yaml holds all the necessary information required to run the scripts in your machine. Make sure to populate this yaml file accordingly.

Note: There is another yaml file called system_prompts.yaml. This is already populated and it holds all the system prompts used in the KG-RAG framework.

Step 5: Run the setup script

Note: Make sure you are in KG_RAG folder

Setup script runs in an interactive fashion.

Running the setup script will:

• create disease vector database for KG-RAG
• download Llama model in your machine (optional, you can skip this and that is totally fine)

python -m kg_rag.run_setup

Step 6: Run KG-RAG from your terminal

Note: Make sure you are in KG_RAG folder

You can run KG-RAG using GPT and Llama model.

Using GPT

# GPT_API_TYPE='azure'
python -m kg_rag.rag_based_generation.GPT.text_generation -g <your favorite gpt model - "gpt-4" or "gpt-35-turbo">
# GPT_API_TYPE='openai'
python -m kg_rag.rag_based_generation.GPT.text_generation -g <your favorite gpt model - "gpt-4" or "gpt-3.5-turbo">

Example:

Note: The following example was run on AWS p3.8xlarge EC2 instance and using KG-RAG v0.3.0.

gpt_demo.mov

Using GPT interactive mode

This allows the user to go over each step of the process in an interactive fashion

# GPT_API_TYPE='azure'
python -m kg_rag.rag_based_generation.GPT.text_generation -i True -g <your favorite gpt model - "gpt-4" or "gpt-35-turbo">
# GPT_API_TYPE='openai'
python -m kg_rag.rag_based_generation.GPT.text_generation -i True -g <your favorite gpt model - "gpt-4" or "gpt-3.5-turbo">

Using Llama

Note: If you haven't downloaded Llama during setup step, then when you run the following, it may take sometime since it will download the model first.

python -m kg_rag.rag_based_generation.Llama.text_generation -m <method-1 or method2, if nothing is mentioned it will take 'method-1'>

Example:

Note: The following example was run on AWS p3.8xlarge EC2 instance and using KG-RAG v0.3.0.

llama_demo.mov

Using Llama interactive mode

This allows the user to go over each step of the process in an interactive fashion

python -m kg_rag.rag_based_generation.Llama.text_generation -i True -m <method-1 or method2, if nothing is mentioned it will take 'method-1'>

Command line arguments for KG-RAG

Argument	Default Value	Definition	Allowed Options	Notes
-g	gpt-35-turbo	GPT model selection	gpt models provided by OpenAI	Use only for GPT models
-i	False	Flag for interactive mode (shows step-by-step)	True or False	Can be used for both GPT and Llama models
-e	False	Flag for showing evidence of association from the graph	True or False	Can be used for both GPT and Llama models
-m	method-1	Which tokenizer method to use	method-1 or method-2. method-1 uses 'AutoTokenizer' and method-2 uses 'LlamaTokenizer' and with an additional 'legacy' flag set to False while initiating the tokenizer	Use only for Llama models

Citation

@article{soman2023biomedical,
  title={Biomedical knowledge graph-enhanced prompt generation for large language models},
  author={Soman, Karthik and Rose, Peter W and Morris, John H and Akbas, Rabia E and Smith, Brett and Peetoom, Braian and Villouta-Reyes, Catalina and Cerono, Gabriel and Shi, Yongmei and Rizk-Jackson, Angela and others},
  journal={arXiv preprint arXiv:2311.17330},
  year={2023}
}