AskMySource
A multimodal RAG system that ingests Git repos, PDFs, and URLs, then answers questions over them with secure on-device LLM inference.
View source on GitHub ↗The problem
Most RAG demos eat one type of source and call it a system. I wanted a single tool that could chew Git repositories, long PDFs, and arbitrary URLs — and answer with sources, locally, without an API key.
The approach
FastAPI for the ingest pipeline, semantic chunking tuned per source-type, FAISS for similarity search, and TinyLlama running locally for the generation step. A Streamlit front-end keeps the UX honest while I iterate on retrieval quality.
The outcome
Useful enough that I now reach for it on every doc-heavy side-project. Working notes: chunking strategy matters more than the model, and a fast UI matters more than either.