MedStudy AI Pro is a high-precision, context-isolated Retrieval-Augmented Generation (RAG) system engineered to ingest dense medical manuscripts and generate hallucination-resistant insights. Optimized for clinical precision, the platform mitigates cross-document context bleeding, optimizes memory utilization, and enforces rigid source grounding.
[User Uploads PDF]
│
▼
[In-Memory Text Extraction] ──► Non-blocking PyPDFLoader stream
│
▼
[Optimized Semantic Chunking] ──► RecursiveCharacterSplitter (1000 tokens | 200 overlap)
│
▼
[HuggingFace Embeddings] ──► Model: all-MiniLM-L6-v2 (384-dim dense vectors)
│
▼
[In-Memory Vector Store] ──► ChromaDB/FAISS (Strict Ephemeral Isolation)
│
▼
[Llama-3.3-70B Pipeline] ──► Temperature: 0.1 + Deterministic Grounding Prompt
│
▼
[Streamlit UI Rendering] ──► Chat Interface + Verified Fragment Citations
🚀 Technical Breakthroughs & Infrastructure During the development of MedStudy AI Pro, several high-impact cloud infrastructure challenges were systematically engineered away:
🔒 Zero-Leakage Ephemeral Storage Standard vector databases persist data to disk, running a risk of context bleeding across separate user files. This engine decouples document tracking by spinning up an isolated, dynamic, in-memory index.
🌐 Bypassing Iframe 403 Rejections Cloud platform dashboards often isolate web frameworks behind <iframe> layers. This project routes user traffic directly to the raw application container network socket.
🧠 Anti-Hallucination Grounding Framework To safely handle delicate clinical data, the inference pipeline utilizes a low-temperature constraint (T = 0.1) coupled with deterministic system instructions.
🛠️ Deep Tech Stack LLM: llama-3.3-70b-versatile via high-speed LPU inference (Groq Cloud). Orchestration: LangChain Core (RunnablePassthrough, ChatPromptTemplate). Vector Embeddings: HuggingFace all-MiniLM-L6-v2. Vector Store: ChromaDB / FAISS (In-Memory Implementation). Frontend: Streamlit Framework (State Handling Session Hooks).
📦 Installation & Developer Setup
-
Prerequisites Python 3.10 or higher. A Groq Cloud API Key.
-
Setup code Bash git clone https://github.com/AYESHAASS/Medical-RAG-Assistant.git cd Medical-RAG-Assistant python -m venv venv source venv/bin/activate # Windows: venv\Scripts\activate pip install -r requirements.txt
-
Environment Create a .env file in the root directory: code Text GROQ_API_KEY=your_production_low_latency_groq_api_token
-
Run code Bash streamlit run app.py
🔬 Rigorous Test Cases Passed The system was evaluated against empirical medical literature to verify retrieval stability across three core stress vectors: Exact Metric Ingestion: Successfully extracted complex statistics (e.g., 95% Confidence Intervals and Brier Scores) from compressed markdown tables.
Multi-Hop Synthesis: Mapped multi-layer neural network parameters (128, 64, 32) across distributed framework logs. Out-of-Bounds Refusals: Corrected handled negative testing by refusing to answer questions not present in the uploaded manuscript.
⚖️ Disclaimer This tool is for educational and research purposes only. It is not intended for clinical decision-making. Always consult primary clinical guidelines and human experts before making medical decisions.
📜 License Distributed under the MIT License. See LICENSE for more information.