KO
|
EN
gitlite — search
Search
#java
#php
#android
#python
#dotfiles
#dotnet
#javascript
#ruby
#cpp
#translation
#map
#c-plus-plus
MDTeamGPT
★ 36
Open GitHub ↗
No description available.
Download README (.md)
Explore Similar Repositories
warpsign
:
A lightning-fast iOS app signing solution that leverages the Apple Developer Portal API for seamless entitlements management and code signing.
DataCommit
:
RAG chatbot for DataCommit podcast episodes with Haystack & Chroma & Gemini
NutriLens-AR-Food-Analysis
:
No description available.
piston
:
Train small sequence models in your browser with WebGPU.
tadoku
:
Blazingly Fast™ Visual Novel tracker
// repository documentation
Was this content helpful?
★ 0
(0 ratings)
Select Rating:
★
★
★
★
★
Submit Feedback
Recent Feedback
×
Download README
Do you want to download the
README.md
file for
MDTeamGPT
?
Download (.md)
# MDTeamGPT 🏥 **A Self-Evolving Multi-Agent Framework for Medical Multi-Disciplinary Team (MDT) Consultations.** MDTeamGPT leverages Large Language Models (LLMs) to simulate a full medical consultation team. It mitigates context collapse using a **Residual Discussion Structure** and employs a **Self-Evolving Mechanism** via dual knowledge bases (CorrectKB & ChainKB) to accumulate medical experience. ## 📅 Latest Update (2025-12-12) * **👁️ Multimodal Support:** Added Vision-Language Model (VLM) integration. Specialists can now analyze uploaded medical images (CT, MRI, X-ray) alongside text descriptions in the first consultation round. * **🔓 Open-Ended QA:** The system now supports and evaluates complex, free-form diagnostic answers (Open-Ended) rather than just multiple-choice questions. * **🛠️ Tool Usage:** Specialists are now equipped with **DuckDuckGo** and **PubMed** tools to autonomously verify facts and retrieve external evidence during reasoning. ## ✨ Key Features * **👨⚕️ Multi-Role Specialists:** Dynamically assigns specialists (e.g., Cardiologist, Neurologist) based on the patient case. * **🧠 Context Engineering:** A **Lead Physician** agent synthesizes discussions into structured residual context (*Consistency, Conflict, Independence, Integration*) to reduce cognitive load. * **🔄 Self-Evolution:** * **CorrectKB:** Stores successful reasoning patterns. * **ChainKB:** Stores reflection and error analysis from incorrect diagnoses. * **🛡️ Safety & Ethics:** A dedicated reviewer ensures consensus convergence and output safety. ## 🚀 Quick Start ### 1\. Prerequisites * Python 3.8+ * OpenAI API Key ### 2\. Installation ```bash git clone https://github.com/KaiChenNJ/MDTeamGPT.git cd MDTeamGPT pip install -r requirements.txt ``` ### 3\. Run the System ```bash streamlit run app.py ``` ## 📂 Project Structure * **`app.py`**: Streamlit UI for case entry, **image upload**, and training visualization. * **`workflow.py`**: LangGraph state machine (Triage $\to$ Consultation $\to$ Safety Check). * **`agents.py`**: Agent definitions including **VLM handling**, **Tool callbacks**, and strict output formatting. * **`knowledge_base.py`**: Dual-memory vector storage (FAISS) for experience retrieval. * **`tools.py`**: **New** integration for Web Search and PubMed tools. ## 🎓 Training Mode 1. **Upload Image** (Optional) and text description. 2. Input the **Ground Truth** (Correct Diagnosis). 3. The system automatically grades the consultation: * **Correct:** Saves reasoning to `CorrectKB`. * **Incorrect:** Performs Chain-of-Thought reflection and saves to `ChainKB`.
