Question-answering with RAG (Retrieval-Augmented Generation) over your Jupyter Notebooks.
Index Jupyter Notebooks and run RAG using LLMs locally:
Llama 2(7B) usingOllamabge-large-enembedding model
Jupyter Notebooks have become de facto a go-to tool for development in Data Science, Machine Learning, and Data Analytics. It seamlessly integrates code, visualizations, and explanatory text, facilitating efficient exploration, analysis, documentation, and sharing of insights
However, as data science, machine learning, and data analysis projects progress, they often accumulate a large number of notebooks, leading to difficulties in usability, management, and addressing questions or queries, creating what is commonly referred to as a "notebook hell" scenario, as illustrated in the example below:
Talk directly to your notebooks with local retrieval-augmented generation (RAG). Index your notebooks and start asking questions. RAG retrieves relevant code and explanations, providing instant answers within your local environment.
Running Large Language Models (LLMs) locally can help avoid sending code or sensitive information to proprietary LLM services by keeping all computations and data processing confined to the local environment, ensuring complete privacy and control.
ollama.ai is used to run Llama 2 locally, see details on installation here.
Setup python environment:
$ PIPENV_VENV_IN_PROJECT=1 pipenv shell
$ pipenv install
Run the streamlit app
$ make run-notebooks-rag
Index your Jupyter Notebooks providing the path to folder.
NOTE: for the testing purposes I used kaggle notebooks (2 pinned notebooks) from the famous Titanic competition. This also explains some of the prompts below.
An example of a prompt for checking which notebooks were indexed:
Some examples of prompts/results:
- Improve chunking, metadata and retrieval
- RAG Evaluation
- Utilize images in RAG
- Index figures
- Embedding on small chunks, retrieve bigger chunks
- Leverage notebooks metadata
- Hybrid Search