Auto-Med AI Agent – Generative Medical Research Assistant

A generative AI agent that reads medical research papers, answers domain-specific questions, summarizes papers, performs PubMed searches, handles calculations, and learns from previous conversations.

---

📌 Project Highlights

• 🧾 Literature-Aware RAG Agent using txt data
• 🔧 Tools Integration: PubMed search, clinical trials search
• 📚 Automatic Question Generation from research papers
• 📏 Evaluation with ROUGE + BLEU
• 💡 Powered by: HuggingFace Transformers, Sentence Transformers, Streamlit, FAISS

---

🏗️ System Architecture

graph TD

A[User Input] --> B{Query Type?}
B -->|Medical Q/A| C[RAG Agent (PDF / Abstracts)]
B -->|clinical trial| D[Clinical trial search tool]
B -->|PubMed| E[PubMed Search Tool]
C --> F[Answer + Summary]
D --> F 
E --> F
G --> A

---

Layer	Tools Used
Frontend	Streamlit
Backend Logic	Python + LangChain-like agent design
NLP Models	Sentence Transformers (MiniLM), HF Pipelines
Tools	Custom Python tools (PubMed,cLINICAL TRIALS)
Evaluation	ROUGE, BLEU

---

git clone https://github.com/mohamed-halemo/Auto-Med-Agent cd automed-ai-agent

Install requirements

pip install -r requirements.txt

Run the app

streamlit run app.py

automed-ai-agent/
│
├── app.py                         # Streamlit frontend
├── agents/
│   ├── literature_agent.py       # Handles RAG with summarization
│   └── tool_agent.py             # Routes queries to tools or literature
│
├── tools/
│   └── toolkit.py                # Calculator, PubMed search
│
├── evaluation/
│   ├── generate_questions.py     # Uses QA pipeline to generate test set
│   └── evaluate_answers.py       # BLEU + ROUGE evaluations
│
├── data/
│   └── generated_qas.json        # Auto-generated Q/A for evaluation
│
├── utils/
│   └── helper.py                 # PDF & text loading utilities
│
├── requirements.txt
└── README.md

How It Works

Upload a PDF or enter a PubMed topic.

The agent creates document embeddings and builds a FAISS retriever.

You can ask:

Domain questions → retrieved + summarized

pubmed cancer → returns top IDs

Memory tracks past Q&A to allow follow-ups.

Metrics + Evaluation Generated 20 Q/A pairs from documents using Hugging Face's transformers pipeline.

Evaluated answers using nltk BLEU and rouge_score metrics.

Used this to benchmark the effectiveness of the current RAG setup.

User query
     ↓
ToolUsingAgent.run()
 ┌─────────────┬─────────────┐
 │ "clinical"            │   "pubmed"    │
 ▼                       ▼
clinical_trial_search()  pubmed_search()
                                             ↓
                                     LiteratureAgent.run()
                                       ┌──────────────┐
                                       │ Semantic     │
                                       │ Search (FAISS│
                                       │ + Bi-Encoder)│
                                       └──────────────┘
                                       ↓
                                   Cross-Encoder Re-rank
                                       ↓
                                  QA Model (extractive)
                                       ↓
                                 Summarization Model
                                       ↓
                              Best Answer + Summary Returned


# By Mohamed Hafez