🧠 Personal Diabetes Assistant — Local RAG System

A lightweight, local Retrieval-Augmented Generation (RAG) system for analyzing diabetes-related patient profiles using LangChain, ChromaDB, and Ollama.
It allows natural language queries over structured CSV medical data, enabling fast and private exploration of patient data — entirely offline.

🚀 How It Works

1. A CSV dataset of patient profiles (medical features like BMI, blood pressure, glucose, etc.) is loaded.
2. Each row is transformed into a textual medical profile (e.g. "55-year-old male with high glucose and low HDL...").
3. These profiles are embedded using the mxbai-embed-large model and stored in a Chroma vector database.
4. When you ask a question, the most relevant profiles are retrieved (based on semantic similarity).
5. A custom prompt is created using those profiles and sent to a local LLM (e.g. llama3) via Ollama.
6. The model generates an informed medical response.

🛠 Installation

1. Clone the repository

git clone https://github.com/mattialoszach/local-rag.git
cd local-rag

2. Install Python dependencies

pip install -r requirements.txt

3. Install Ollama and the required models

ollama pull llama3
ollama pull mxbai-embed-large

📄 Dataset

The system expects a CSV file named diabetes_dataset.csv in the root directory, with the following columns:

id, Age, Sex, Ethnicity, BMI, Waist_Circumference, Fasting_Blood_Glucose, HbA1c, Blood_Pressure_Systolic, Blood_Pressure_Diastolic, Cholesterol_Total, Cholesterol_HDL, Cholesterol_LDL, GGT, Serum_Urate, Physical_Activity_Level, Dietary_Intake_Calories, Alcohol_Consumption, Smoking_Status, Family_History_of_Diabetes, Previous_Gestational_Diabetes

You can replace the dataset with your own medical records (same format). Dataset can be found here: Kaggle

💬 Usage

python main.py

Then simply type natural language questions, such as:

• “Show patients with high fasting glucose and low HDL.”
• “Which profiles have signs of metabolic syndrome?”
• “Find people with a family history of diabetes and obesity.”

Type '/q', '/quit' or '/exit' to quit.

🎯 Use Case

This tool is ideal for:

• Local clinical research and education
• Medical case exploration and risk group analysis
• Privacy-preserving data querying without cloud APIs
• Prototyping AI-based assistants for structured health data

🧩 Possible Extensions

• Add a web interface or GUI
• Support for PDF or JSON patient records
• Integrate a diagnosis suggestion model
• Include follow-up questions (multi-turn RAG)
• Add further API support

🔒 Fully Local and Offline

All processing is done locally:

• No cloud APIs
• All data stays on your machine
• Perfect for sensitive health-related use

⚠️ This project is for educational and research purposes only. Not intended for medical diagnostics.