MQNotebook — Enterprise-Grade RAG System

Overview

MQNotebook is a production-ready, local-first Retrieval-Augmented Generation (RAG) system built to handle the messy reality of enterprise documents.

Unlike typical “chat with PDF” demos, MQNotebook is engineered to ingest:

• scanned PDFs,
• complex Excel spreadsheets,
• PowerPoint decks with speaker notes,
• and mixed-format document collections,

while running securely on a local machine or cloud environment.

🔗 Live Demo: https://mqnotebook.streamlit.app/

Why MQNotebook Exists

Most Retrieval-Augmented Generation (RAG) demos assume an ideal world: clean PDFs, perfect text extraction, and well-structured documents.

Enterprise reality is different.

In real environments, knowledge lives in:

• scanned documents
• poorly formatted PDFs
• spreadsheets with implicit structure
• slide decks where critical context lives in speaker notes

Most RAG pipelines fail before retrieval even begins — at ingestion.

MQNotebook exists to address this gap.

Instead of optimizing for model size or prompt tricks, MQNotebook focuses on:

• failure-first document ingestion
• retrieval precision over raw recall
• controlled context passed to the LLM

The system is designed so that the language model only sees high-confidence, conceptually relevant information, reducing hallucinations and wasted tokens.

Design Philosophy

Most RAG demos assume clean text input.
MQNotebook assumes documents will fail — and designs for that failure.

The system prioritizes:

• deterministic ingestion,
• retrieval precision,
• and controlled LLM exposure.

System Architecture

MQNotebook follows a multi-stage retrieval pipeline that separates document ingestion, retrieval, and reasoning.

Document Ingestion Layer

• Scanned PDFs & Images
  OCR fallback using Tesseract + Poppler when text extraction fails.

• PowerPoint (.pptx)
  Extracts text from slides, shapes, SmartArt, and speaker notes.

• Excel (.xlsx)
  Parses spreadsheets row-by-row while preserving column structure.

All extracted content is normalized before embedding.

Retrieval & Ranking Layer

• Vector Store: ChromaDB
• Embeddings: BAAI/bge-small-en-v1.5 (runs locally on CPU)
• Retriever: Top-K semantic search
• Reranker: Cross-encoder to filter results before LLM inference

This reduces hallucinations by ensuring only conceptually relevant context reaches the model.

Generation Layer

• LLM: Gemini 2.0 Flash (via OpenRouter)
• Large context window
• Fast reasoning
• Controlled prompt scope

API keys are handled in-memory per session and cleared on reset.

Engineering Considerations

• Windows File Locking Fix
  Dynamic, timestamped sessions prevent WinError 32 during vector store resets.

• OS-Aware Execution
  Automatically switches OCR binary paths between Windows (local) and Linux (cloud).

• Local-First Security Model
  Documents and embeddings remain local; only final prompts are sent to the LLM.

Relationship to DevShelf

MQNotebook complements DevShelf, which focuses on classic information retrieval and search engines.

• DevShelf: First-principles search engine (TF-IDF, inverted index, ranking)
• MQNotebook: Modern RAG system (embeddings, reranking, LLM integration)

Together, they demonstrate a full spectrum of search and knowledge retrieval systems — from traditional IR to AI-augmented reasoning.

Source & Documentation

• Source Code: https://github.com/Kas-sim/MQNotebook
• DevShelf: https://github.com/Kas-sim/DevShelf