This is the GitHub repository for our IEEE VIS 2024 paper,
“KNOWNET: Guided Health Information Seeking from LLMs via Knowledge Graph Integration” (Pre-Print).
This paper received the Best Paper Honorable Mention at IEEE VIS 2024.

Demo: https://maurilaparva.github.io/KNOWNET/

Documentation: https://visual-intelligence-umn.github.io/KNOWNET/

git clone https://github.com/visual-intelligence-umn/KNOWNET.git
cd KNOWNET

python3 -m venv venv
source venv/bin/activate   # (on Windows: venv\Scripts\activate)

KNOWNET uses React + TypeScript + Vite.

pnpm install

(If you don’t have pnpm installed, run npm install -g pnpm first.)

pnpm run dev

The app will be served locally at:
http://localhost:5173

If you are modifying the codebase:

pnpm install
pnpm run dev

cd api
python3 index.py

Both servers must run concurrently in development mode.

The KNOWNET backend relies on two primary data resources:

1. Embedding Files — Precomputed text embeddings used for entity and relation verification.
   These embeddings are stored on the server and accessed through embedding_utils.py and embeds.py.
   Developers reproducing the system locally can use existing OpenAI or local embedding models to regenerate these vectors from the source entity/relation files.

2. Knowledge Graph Database (Neo4j) —
   The deployed version of KNOWNET connects to a Neo4j database hosted on AWS EC2, which contains the curated biomedical knowledge graph used for verification and recommendation.
   The hosted instance supports the public demo and is read-only for external users.

KNOWNET uses two different Neo4j environments, each serving a different purpose:

• Hosts the curated, final biomedical knowledge graph used in the VIS 2024 paper
• Contains data imported from node_data.csv and rel_data_filtered.csv, plus several rounds of filtering and manual verification
• Read-only for all external users
• Connected directly to the deployed Flask backend
• Includes precomputed embeddings and optimized indexes
• Intended only for running the public demo, not for developer experimentation

• Used when developers want to:
  • extend KNOWNET
  • test custom knowledge graphs
  • run experiments offline
• Requires creating your own copies of node_data.csv and rel_data_filtered.csv based of ADInt_CUI_embeddings.parquet (can be downloaded below).
• Requires computing or providing your own embeddings (see next section)
• Credentials and database URI must be configured locally
• Fully editable and meant for development

To reproduce or extend KNOWNET locally with a custom knowledge graph:

1. Install Neo4j Community or AuraDB

   • Local setup: https://neo4j.com/download/
   • Cloud setup (recommended for collaboration): https://neo4j.com/cloud/aura/

2. Create a New Database

   • Configure a database (e.g., knownet_db) and note your connection credentials:

     NEO4J_URI=bolt://localhost:7687
     NEO4J_USER=neo4j
     NEO4J_PASSWORD=<your-password>
     

   • Update these credentials in your local .env or configuration file (used by verify.py and recommend.py).

3. Load Graph Data The biomedical knowledge graph used for KNOWNET is constructed from two ADInt-derived files:

   • node_data.csv - list of biomedical entities
   • rel_data_filtered.csv - filtered relations between entities These files were used to build both the AWS Neo4j database and the local development setup. Note: KNOWNET does not include these files in the repository.

      LOAD CSV WITH HEADERS FROM 'file:///node_data.csv' AS row
      CREATE (:Entity {
      cui: row.CUI,
      name: row.name,
      type: row.type
      });
     
      LOAD CSV WITH HEADERS FROM 'file:///rel_data_filtered.csv' AS row
      MATCH (a:Entity {cui: row.Source}), (b:Entity {cui: row.Target})
      CREATE (a)-[:RELATION {type: row.Relation}]->(b);
     

   • If you are using your own biomedical dataset, you may adapt this format to match your schema.

4. Generate or Import Embeddings

   • Use embedding_utils.py to compute embeddings for all entities and relations:

     python3 api/embedding_utils.py

   • Store resulting vectors locally or in a connected S3 bucket.

KNOWNET uses precomputed concept embeddings stored in:

• ADInt_CUI_embeddings.parquet

A copy of the embeddings file can be downloaded here:

Download ADInt_CUI_embeddings.parquet

After downloading the file, place it in the api/ directory (or your preferred location), and update the path in embeds.py:

EMBEDDING_FILE = "api/ADInt_CUI_embeddings.parquet"

The hosted KNOWNET backend runs on AWS EC2 with a connected Neo4j instance configured for read access.
Developers replicating this deployment can follow a similar setup:

• Launch an EC2 instance (Ubuntu 22.04 or later).
• Install Neo4j, Flask, and dependencies listed in requirements.txt.
• Configure security groups to allow bolt:// (7687) and HTTPS (443) traffic.
• Use Nginx or a similar reverse proxy for routing and SSL termination.

The KNOWNET repository follows a modular full-stack architecture composed of a Python-based backend and a React + TypeScript frontend.
Below is an overview of the top-level structure:

KNOWNET/
├── api/                     # Flask backend (embedding, verification, recommendation)
│   ├── embedding_utils.py
│   ├── embeds.py
│   ├── index.py
│   ├── recommend.py
│   └── verify.py
│
├── src/                     # React + TypeScript frontend
│   ├── assets/
│   ├── components/
│   ├── lib/
│   ├── main.tsx
│   ├── App.css
│   └── index.css
│
├── docs/                    # Documentation and setup guides
├── requirements.txt         # Python dependencies
├── package.json             # Frontend dependencies
├── vite.config.js           # Vite build configuration
├── tailwind.config.js       # TailwindCSS configuration
├── postcss.config.js        # PostCSS configuration
├── run_flask.sh             # Flask startup script
└── README.md

This directory contains the primary interface modules that enable chat-based interaction, visualization, and dynamic user feedback.

If you use or reference this project, please cite:

Youfu Yan, Yu Hou, Yongkang Xiao, Rui Zhang, and Qianwen Wang. 2025.
KNowNet: Guided Health Information Seeking from LLMs via Knowledge Graph Integration.
IEEE Transactions on Visualization and Computer Graphics, 31(1), 547–557.
https://doi.org/10.1109/TVCG.2024.3456364

A follow-up version of KNOWNET explores web search–based validation as an alternative to static knowledge-graph integration.
This prototype introduces edge-level uncertainty quantification to assess the reliability of model-generated relations in real time.

View KNOWNET v2 here: https://maurilaparva.github.io/prototype/