Doc Sage

A document Q&A assistant that uses RAG (Retrieval-Augmented Generation) to answer questions about PDF documents with page-level citations.

Features

• PDF processing with page-level chunking
• Semantic search using cosine similarity (threshold: 0.3)
• Streaming responses with conversation history
• Precise page citations for transparency

Requirements

• Python 3.13+
• LM Studio running on http://localhost:1234
• Models: text-embedding-nomic-embed-text-v1.5, qwen/qwen3-4b

Quick Start

# Install dependencies
uv sync

# Or with pip
pip install lmstudio scikit-learn pymupdf

from main import RAG

rag = RAG("document.pdf")
rag.chat("What are the main findings?")  # Streams response to console

How it Works

1. Extract - Splits PDF into page-based chunks with metadata
2. Embed - Creates embeddings for query and document pages
3. Retrieve - Finds top 7 most similar chunks (cosine similarity ≥ 0.3)
4. Generate - Uses LM Studio to create contextual responses with page citations