Local PDF Chatbot

Local-first PDF question answering app for indexing a single document, storing embeddings locally, and answering only from the uploaded content.

This project is built with Streamlit, ChromaDB, SentenceTransformers, PyMuPDF, and Ollama. A user uploads one PDF, the app extracts and chunks the text, stores embeddings in a persistent local vector database, and answers questions using only the uploaded document.

Features

• Upload and index a PDF through a Streamlit web interface
• Extract text locally with PyMuPDF
• Chunk content automatically with LangChain text splitters using fixed app defaults
• Create embeddings locally with all-MiniLM-L6-v2
• Persist vectors locally with ChromaDB
• Generate answers locally with an Ollama model such as llama3 or mistral
• Return the exact fallback message when the answer is not supported by the document
• Let the model synthesize summaries and answers from the retrieved document chunks

Project Structure

app.py
config.py
ingest.py
qa.py
vector_store.py
tests/
report/
presentation/

Requirements

• Python 3.11 or 3.12
• uv for environment and dependency management
• Ollama installed locally
• An Ollama model pulled locally, for example llama3

Setup

Create a fresh environment with a supported Python version:

uv venv --python 3.12
source .venv/bin/activate
uv sync

Pull a local Ollama model if you do not already have one:

ollama pull llama3

Run the App

source .venv/bin/activate
streamlit run app.py

Then open the local Streamlit URL, upload a PDF, click Index PDF, and start asking questions. Chunk size and overlap are handled automatically by the app.

How It Works

1. The user uploads a PDF in the Streamlit sidebar.
2. PyMuPDF extracts text page by page.
3. LangChain smart chunking breaks each page on likely headings, paragraphs, bullets, and sentence boundaries before falling back to smaller splits.
4. SentenceTransformers creates embeddings for each chunk.
5. ChromaDB stores the embeddings and metadata locally.
6. The user question is embedded and matched against the indexed chunks.
7. Retrieved chunks are sent to Ollama with a synthesis prompt that can answer questions or summarize the document.
8. If support is not found, the app returns:

The information is not available in the uploaded document.

Notes

• The first version is optimized for one small PDF at a time.
• The vector collection is reset whenever a new PDF is indexed.
• The UI does not yet show citations, but page metadata is stored for future extension.

Suggested Demo Flow

1. Upload a short PDF of 5-6 pages.
2. Ask one direct factual question from the document.
3. Ask a paraphrased version of the same question.
4. Ask an unsupported question to show the fallback behavior.