fd

All work

DocuMind AI

Shipped

A local AI assistant that answers questions about your PDFs, with page citations, and never sends your documents anywhere.

Stack
Python, LangChain, FAISS, Ollama (Llama 3, Mistral, LLaVA), Streamlit
Links
DocuMind’s PDF Q&A tab. An answer to “What is this document about?” sits above an open Sources panel reading: [1] text.pdf, p. 1, followed by the quoted passage.
Every answer can be checked. Open Sources and citation [1] names the file and page, text.pdf, page 1, and quotes the exact passage the answer came from.Full size

What it does

Upload a PDF, ask a question in plain language, and get an answer with numbered citations. Open Sources under any answer and you see the file, the page, and the quoted text each citation came from, so an answer can always be checked against the document itself.

Why it runs locally

There is no hosted LLM API anywhere in the pipeline. The chat models (Llama 3, Mistral, and LLaVA for images) run on your own machine through Ollama, the embeddings come from nomic-embed-text-v2-moe, and the vector index lives in FAISS. Your documents never leave the computer they’re on.

Built as a capstone

DocuMind is a retrieval-augmented generation (RAG) pipeline built from scratch as a university capstone, with its own evaluation set.

Proof

ClaimHow it was checkedState
294 of 294 tests passpytestVerified 27 Sep 2026
Embeddings come from nomic-embed-text-v2-moedocumind/config.py in the repoVerified 27 Sep 2026

Screens

DocuMind’s Chat tab with LLaVA selected as the chat model and an image upload in the sidebar.
The Chat tab. Plain chat, or upload an image in the sidebar and ask about it.Full size
The PDF Q&A tab with an empty Upload PDFs box.
The PDF Q&A tab, before any document is added.Full size
text.pdf uploaded, with a spinner reading “Reading and indexing text.pdf”.
Reading and indexing the uploaded PDF.Full size
A confirmation that text.pdf is indexed, 1 page and 1 chunk in 4.3 seconds, with “Answering from text.pdf” below it.
Indexed. One page, one chunk, 4.3 seconds.Full size
An answer about the document’s project budget and report deadline that ends with the citation [1], and a closed Sources panel under it.
The answer, ending in its [1] citation.Full size