|
--- |
|
title: Med |
|
emoji: 🏆 |
|
colorFrom: blue |
|
colorTo: purple |
|
sdk: streamlit |
|
sdk_version: 1.43.2 |
|
app_file: app.py |
|
pinned: false |
|
short_description: Medical field with next‑gen technology |
|
--- |
|
|
|
|
|
# AI-Powered Medical Knowledge Graph Assistant (Mini-LADDER Demo) |
|
|
|
This repository demonstrates a Streamlit application that: |
|
- Retrieves PubMed abstracts via NCBI’s E-utilities. |
|
- Indexes and retrieves relevant documents using ChromaDB. |
|
- Generates biomedical answers using Microsoft BioGPT-Large-PubMedQA. |
|
- Applies a two-stage self-improvement mechanism (inspired by Tufa Labs’ LADDER) that: |
|
- Generates naive sub-questions. |
|
- Produces an initial answer. |
|
- Self-critiques and refines the answer. |
|
- Visualizes key terms in a knowledge graph using PyVis. |
|
|
|
## Key Features |
|
|
|
1. **PubMed + Chroma**: Retrieves and indexes relevant abstracts. |
|
2. **BioGPT-Large-PubMedQA**: Generates an initial answer. |
|
3. **Mini-LADDER Approach**: |
|
- **Sub-Question Decomposition**: Generates sub-questions from the main query. |
|
- **Self-Critique & Refinement**: Uses a second pass to critique and refine the answer. |
|
4. **Interactive Knowledge Graph**: Displays a PyVis graph of the top documents and key terms. |
|
|
|
## Setup Instructions |
|
|
|
1. **Install Dependencies** |
|
|
|
Create a virtual environment (optional) and install the required packages: |
|
```bash |
|
pip install -r requirements.txt |
|
Set Environment Variables |
|
|
|
Set your PubMed API key: |
|
|
|
bash |
|
Copy |
|
export PUBMED_API_KEY=<YOUR_NCBI_API_KEY> |
|
Run the App |
|
|
|
Launch the Streamlit app: |
|
|
|
bash |
|
Copy |
|
streamlit run app.py |
|
Access the App |
|
|
|
Open your browser at http://localhost:8501. |
|
|
|
Project Structure |
|
Copy |
|
. |
|
├── app.py |
|
├── backend.py |
|
├── mini_ladder.py |
|
├── retrieval.py |
|
├── visualization.py |
|
├── README.md |
|
└── requirements.txt |
|
About the Mini-LADDER Approach |
|
Inspired by Tufa Labs’ LADDER (Learning through Autonomous Difficulty-Driven Example Recursion), this demo shows how a model might: |
|
|
|
Decompose a query into simpler sub-questions. |
|
Generate an initial answer using retrieval-augmented generation. |
|
Self-critique and refine the answer based on detected gaps. |
|
Ultimately, this approach hints at how autonomous, recursive learning could be implemented. |
|
Enjoy exploring potential extensions into code generation, theorem proving, or other domains! |
|
|
|
yaml |
|
Copy |
|
|
|
--- |
|
|
|
## File: `requirements.txt` |
|
|
|
```txt |
|
streamlit |
|
pyvis |
|
chromadb |
|
transformers |
|
sentence-transformers |
|
torch |
|
requests |
