An MLHub Package for Personalized Health Queries

A simple toolkit to work with personal health data and LLMs locally.

This MLHub package is designed to create a Retrieval-Augmented Generation (RAG) system for personalized healthcare queries using large language models (LLMs) and private health data. The system operates locally to ensure privacy is preserved.

Overview

The system consists of several components working together and makes use of public libraries (such as LangChain to orchestrate the retrieval and integration of private health data and Ollama as the underlying large language model (LLM) provider) to enable personalized health queries:

1. Document Ingestion: Documents (such as lab reports or health records) are loaded and processed into smaller chunks. Some supported formats include PDF, TXT, DOCX, and MD.

2. Vector Store Creation: The chunks of text are then converted into embeddings using an LLM, such as nomic-embed-text, and stored in a FAISS vector store. This allows for fast similarity-based searches on the data.

3. Query Execution: When a user submits a query, the system uses the vector store to find relevant information and generate a response. The query can be processed using RAG techniques, combining the retrieved information with the language model’s generation capabilities.

This system ensures that healthcare queries can be answered using personal data without compromising privacy, as all computations happen locally.

Quick Start

First, you'll need to install the mlhub package and health_rag, then configure them. This process may take some time depending on your system and internet speed.

# Install and configure mlhub
pip3 install mlhub
ml configure

# Install and configure health_rag
ml install ar4152/health_rag@main
ml configure health_rag

Now, run a simple example. This should execute quickly.

# Download sample health data
wget https://www.testing.com/wp-content/uploads/2021/07/CBC-sample-report-with-notes_0.pdf -O sample_lab_report.pdf

# Create a context with the downloaded report
ml create_context health_rag sample_lab_report.pdf

# Ask a query related to the report
ml query health_rag "What is my RBC count? Is it normal?" --vectorstore_path ./vector_store

Commands Overview

1. chat_mode Start an interactive chat with memory.

   ml chat_mode health_rag
   

2. query Answer health questions with optional retrieval from your data.

   ml query health_rag "Explain my lab results." --vectorstore_path vector_store/
   

3. pull_model Download LLM models for offline use.

   ml pull_model health_rag llama3.2:1b
   

4. remove_model Remove downloaded models to free space.

   ml remove_model health_rag llama3.2:1b
   

5. create_context Build a FAISS vector store from documents.

   ml create_context health_rag ~/Documents/HealthData --save_dir vector_store
   

For full usage and examples, see the detailed commands documentation.

Default Models

• Embedding Model: The default model for generating embeddings from the document chunks is nomic-embed-text. You can replace it with any other model if preferred by using the --embedding_model flag when creating the context.

• Query Model: By default, queries are processed using llama3.2:1b.

License

This code is licensed under the MIT License.