Info

GraphMind GraphMind

GraphMind

12/21/2024

GraphMind AI Assistant

GraphMind is an advanced AI research assistant leveraging LangGraph for agent orchestration and Neo4j to construct and query a knowledge graph of research papers. This repository contains the codebase and setup instructions for deploying GraphMind locally.

Table of Contents

  1. Features
  2. GraphMind Agent Architecture
  3. Requirements
  4. Clone the Repository
  5. Container Setup
  6. Setup Instructions
  7. Running the Application
  8. Future Updates
  9. References

Features

  • Constructs a knowledge graph from a corpus of documents.
  • Provides agentic planning & reasoning using LangGraph, based on the Graphreader paper.
  • Agent will traverse knowledge graph database and update its “notebook” with relevant information to answer user queries.
  • Fully containerized with Docker for on-premises hosting.
  • LLM serving via Ollama (currently OpenAI API), with predefined structured output.

GraphMind Agent Architecture

The underlying architecture behind the GraphMind Agentic Framework leverages the Graphreader paper. The diagrams below show the proposed implementation:

GraphMind Proposed ArchitectureSemantic Router AgentResearch SubgraphAgent Node Architecture

The Neo4j Knowledge Graph ontology is pre-defined and shown as follows:

Neo4j Ontology

Requirements

Ensure the following are installed on your system:

Clone the Repository

Start with cloning the repository into your local drive.

Terminal window
git clone https://github.com/michael-hoon/graphreader-agent.git
cd graphreader-agent

Container Setup

Note: Current version does not have docker-compose setup yet, have to manually setup containers for now. Requires Neo4j and MinIO S3 images from Docker hub.

Neo4j Container

Start up another terminal instance. Pull the official Neo4j image from the docker hub repository:

Terminal window
docker pull neo4j

Start an instance of the database:

Terminal window
docker run \
    --publish=7474:7474 --publish=7687:7687 \
    --volume=$HOME/neo4j/data:/data \
    neo4j

Visit the WebUI host on http://localhost:7474/browser/, and login with the default username/password: neo4j:neo4j. The UI will prompt you to set a different password. Save this for use in the .env file later.

MinIO Container

Run the following command (on another instance of your terminal) to start a container instance of the MinIO S3 object storage (automatically pulls the latest image). Remember to change the installation path of the cloned repository.

Terminal window
docker run \
    -p 9000:9000 \
    -p 9001:9001 \
    --name graphreaderminio \
    -v /path-to-your-installation/graphreader-agent/docs:/pdfdata \
    quay.io/minio/minio server /pdfdata --console-address ":9001"

Visit the WebUI host on http://127.0.0.1:9001/ and login with the default credentials minioadmin:minioadmin. Change your username/password if required, and remember these details for the .env file later.

S3 Bucket Creation

Create a bucket and name it appropriately (e.g. graphreader-docs), without any of the features selected. You should see an option to ‘Browse Files’, select it and upload the two sample PDF files from /graphreader-agent/docs/ or your own personal PDF files.

(Optional) Verify Containers

Verify that all relevant containers are running:

Terminal window
docker ps

Setup Instructions

1. Set Up Virtual Environment

On the first terminal instance where you cloned the project repository, create a virtual environment in the graphreader-agent directory with uv and activate it:

Terminal window
pip install uv # ignore if uv is already installed on system
uv sync --frozen
# "uv sync" creates .venv automatically


source .venv/bin/activate

2. Configure the .env file

Copy the .env.example file and customize it:

Terminal window
cp .env.example .env

Edit the file accordingly. You will require your OpenAI API key, Neo4j database and MinIO S3 storage details from before.

Running the Application

  1. Start the Streamlit frontend:
Terminal window
cd /src/streamlit_app
streamlit run app.py

  1. Access the app in your browser at http://localhost:8501.

  2. On the sidebar, select the “Process Knowledge Graph” option to run the agent and process your PDF files into the Neo4j database. You can now go to the Neo4j Web console to see the resulting Knowledge Graph.

✨ Interact with the GraphMind assistant and analyze your PDF research corpus. Ask any question!

✨ Give feedback on LLM answers and export conversation history on the “Chat Management” tab in JSON format, with human feedback tracked.

Streamlit DemoNeo4j Example

Known Issues / Future Updates

  1. Currently no functionality for choosing different models on sidebar (config file WIP).
  2. Ollama structured output still having issues, using OpenAI API for now.
  3. PostgreSQL integration failing due to issue with tracking session states.
  4. Agent implementation needs improvement by rewriting certain agents as tools instead.
  5. Agent “thoughts” not yet configured properly in streamlit containers, rewriting agents as tools might help. Rerunning streamlit instance removes the container too.
  6. FastAPI not integrated yet for Multi-user support on deployed platform
  7. Containerisation not configured for graphreader-agent code.
  8. Container management with Kubernetes not yet set up.
  9. Monitoring/Observability platforms (Grafana) not set up as well.

References

← Back to projects