A pragmatic experiment in chatting with PDFs locally

What it does

This is a retrieval-augmented generation chatbot that ingests multiple PDFs and answers questions based on their contents. It wraps the pipeline in a simple Gradio web interface and targets fully local execution via Ollama, with Hugging Face models available as an alternative backend. You can run it on your own machine or inside a Kaggle notebook tunneled through Ngrok.

The interesting bit

The README labels the whole thing an “experiment” in wiring together existing parts—document retrieval, local inference, and a lightweight UI—rather than claiming to invent new retrieval algorithms. That transparency is welcome; the project is essentially a pragmatic reference implementation for developers who want to see how the pipes fit without reading separate docs for each component.

Key highlights

• Runs fully offline through Ollama, with Hugging Face models as an alternative backend.
• Accepts multiple PDF inputs in a single session.
• Browser-based UI via Gradio.
• Kaggle notebook path for users who want to experiment before committing to a full local install.
• Multilingual chat is listed as a coming-soon feature.

Caveats

• Multilingual support is marked “Coming soon,” so the interface appears to be English-first for now.
• The README is heavy on setup scripts and environment provisioning, but light on architecture details; the RAG flow diagrams are shown without much explanation of what differentiates this specific pipeline.
• It is fundamentally glue code connecting a UI to local inference backends—useful, but not a novel framework.

Verdict

Worth a look if you need a quick, local PDF Q&A prototype and prefer to assemble proven components rather than build from scratch. Skip it if you are after rigorous citation accuracy or a mature multilingual pipeline.