Skip to content
BranchesTags

Repository files navigation

DocuMind AI-Powered PDF Chat Assistant

DocuMind Banner

Python Flask Google Gemini FAISS License: MIT

Read PDFs and ask questions — without leaving the page.

📖 What is DocuMind?

DocuMind is a Retrieval-Augmented Generation (RAG) web application that lets you read a PDF document on the left and chat with an AI about it on the right — all in one browser tab.

No tab-switching. No copy-pasting. Just open a document and ask.

Upload a research paper, textbook, contract, or technical manual — then ask questions, get summaries, work through equations, or request explanations in plain language. The AI answers using only the content of your document.

✨ Key Features

Feature Details
📄 PDF Viewer Embedded side-by-side PDF reader with drag-and-drop upload
🤖 RAG Chat Hybrid BM25 + vector search feeds precise context to the LLM
🧮 Math Rendering Full KaTeX support renders $inline$ and $$block$$ LaTeX
📝 Notes Save, edit, download any Q&A pair as a personal note
🌙 Dark / Light Mode Persisted theme toggle
⚡ PDF Caching First upload processes and caches; subsequent opens are instant
🖼️ Image OCR EasyOCR + CLIP for on-demand image understanding
🔧 Resizable Panes Drag the divider to resize the PDF and chat panels

🏗️ System Architecture

flowchart TD
    subgraph Browser["🌐 Browser (Frontend)"]
        UI["DocuMind UI\nHTML + CSS + JS"]
        PDF["PDF Viewer\n(iframe embed)"]
        Chat["Chat Panel\nMarked.js + KaTeX"]
        Notes["Notes Panel\nlocalStorage"]
    end

    subgraph Flask["🐍 Flask Server (app.py)"]
        Upload["POST /upload"]
        Query["POST /query"]
        Serve["GET /uploaded/<file>"]
    end

    subgraph RAG["🔍 RAG Pipeline (rag.py)"]
        Extract["PDF Extractor\nPyMuPDF"]
        Chunk["Text Chunker\nLangChain CharacterTextSplitter"]
        Embed["Sentence Embedder\nBAAI/bge-small-en-v1.5"]
        Index["Vector Index\nFAISS IndexFlatL2"]
        Cache["Cache Layer\nMD5 Hash + Pickle"]
        Hybrid["Hybrid Search\nBM25 + Vector (α=0.5)"]
        Clip["Image Analysis\nCLIP + EasyOCR"]
    end

    subgraph LLM["🤖 Google AI"]
        Gemini["Gemma 4 31B IT\ngemma-4-31b-it"]
    end

    UI -- "Upload PDF" --> Upload
    UI -- "Ask question" --> Query
    Upload -- "save + process" --> Extract
    Extract --> Chunk --> Embed --> Index
    Embed --> Cache
    Index --> Cache
    Query -- "user query" --> Hybrid
    Hybrid -- "top-k chunks" --> Gemini
    Hybrid -- "image reference?" --> Clip
    Clip --> Gemini
    Gemini -- "answer" --> Query
    Query -- "JSON response" --> Chat
    Serve --> PDF
Loading

Data Flow Query Time

sequenceDiagram
    participant U as User
    participant F as Frontend
    participant S as Flask
    participant R as RAG Engine
    participant G as Gemini API

    U->>F: Types question + hits Enter
    F->>S: POST /query { query }
    S->>R: handle_query(query)
    R->>R: hybrid_search()<br/>BM25 + FAISS top-5 chunks
    R->>G: generate_content(prompt + context)
    G-->>R: answer text
    R-->>S: { response, chat_history }
    S-->>F: JSON { response }
    F->>F: Marked.js → markdown<br/>KaTeX → math
    F-->>U: Rendered AI response
Loading

PDF Processing Pipeline

flowchart LR
    A[📄 PDF Upload] --> B[PyMuPDF\nExtract text + images]
    B --> C{Cache\nexists?}
    C -- Yes --> D[Load .pkl cache]
    C -- No --> E[CharacterTextSplitter\nchunk_size=500, overlap=50]
    E --> F[BAAI/bge-small-en-v1.5\nEncode chunks]
    F --> G[FAISS IndexFlatL2\nStore embeddings]
    G --> H[Save .pkl cache]
    D --> I[🟢 Ready to query]
    H --> I
Loading

🛠️ Tech Stack

Backend

Library Role
Flask Web server & REST API
PyMuPDF (fitz) PDF text and image extraction
LangChain CharacterTextSplitter for chunking
sentence-transformers BAAI/bge-small-en-v1.5 dense embeddings
FAISS Approximate nearest-neighbour vector search
rank-bm25 Sparse BM25 keyword retrieval
NLTK Tokenisation for BM25
CLIP Image-text similarity scoring
EasyOCR On-demand OCR for PDF images
google-genai Google Gemini API client (gemma-4-31b-it)
python-dotenv Environment variable management

Frontend

Technology Role
Vanilla JS App logic, resizer, tabs, upload, chat
Marked.js Client-side Markdown rendering
KaTeX LaTeX math rendering ($...$, $$...$$)
DOMPurify XSS sanitisation of AI output
CSS Custom Properties Design tokens for full dark/light theming
Inter + JetBrains Mono Google Fonts typography

🚀 Setup & Installation

Prerequisites

Requirement Notes
Python 3.10+ python.org
Tesseract OCR UB Mannheim installer (Windows)
Google AI API Key Get one at ai.google.dev
Git For cloning the repo

1. Clone the repository

git clone https://github.com/avish006/Rag-Project.git
cd Rag-Project

2. Create and activate a virtual environment

# Windows
python -m venv .venv
.venv\Scripts\activate

# macOS / Linux
python -m venv .venv
source .venv/bin/activate

3. Install dependencies

pip install -r requirements.txt

⚠️ First install is large (~3–5 GB) — PyTorch, CLIP, sentence-transformers, and EasyOCR models will download on first run.

4. Download NLTK data

python nltk_download.py

5. Configure environment variables (Optional)

Create a .env file in the project root:

FLASK_SECRET_KEY="any-random-secret-string"
# TESSERACT_CMD="C:\\Program Files\\Tesseract-OCR\\tesseract.exe" (Windows only)

🔑 API Keys: You no longer need to put your Google API key in the .env file. The app now uses a Bring Your Own Key (BYOK) model. You will enter your key securely directly in the browser UI when you launch the app. ⚠️ Never commit your .env file — it is listed in .gitignore.

6. Run the application

python app.py

Then open http://127.0.0.1:5000 in your browser.

📝 First startup takes 15–30 seconds — PyTorch and model weights load into memory. Subsequent restarts are faster once models are cached locally.

📁 Project Structure

Rag-Project/
├── app.py                  # Flask server & REST API routes
├── rag.py                  # RAG pipeline (extract, embed, retrieve, generate)
├── requirements.txt        # Python dependencies
├── nltk_download.py        # NLTK corpus downloader script
├── .env.example            # Template for environment variables
├── .gitignore
│
├── templates/
│   └── index.html          # Single-page app shell
│
├── static/
│   ├── style.css           # Full dark/light design system (CSS custom properties)
│   ├── script.js           # Frontend logic (chat, upload, resizer, notes, math)
│   └── marked.min.js       # Bundled Markdown parser
│
├── uploads/                # Uploaded PDFs (gitignored)
├── cache/                  # Pickle cache of processed PDFs (gitignored)
├── figures/                # Extracted PDF images (gitignored)
└── model_cache/            # Downloaded model weights (gitignored)

💡 Usage Guide

Uploading a PDF

  1. Drag and drop any PDF onto the left panel or click Choose PDF
  2. Wait for the progress bar — processing time depends on PDF size
  3. The PDF opens in the viewer once processing completes

Asking Questions

  • Type your question in the chat box on the right
  • Press Enter to send, Shift+Enter for a new line
  • Use the suggestion chips on the empty screen for quick starts

Math in Responses

The AI can render full LaTeX math. Examples you can ask:

  • "Derive the quadratic formula"
  • "What does this equation mean: $E = mc^2$?"
  • "Show me the integral for the normal distribution"

Saving Notes

  • Click Save to Notes on any AI response
  • Give the note a name in the modal
  • Switch to the Notes tab to view, edit, or export your notes

⚙️ Configuration

Variable Default Description
FLASK_SECRET_KEY — Flask session secret (set in .env)
TESSERACT_CMD C:\Program Files\Tesseract-OCR\tesseract.exe Path to Tesseract binary
MAX_CONTENT_LENGTH_MB 50 Max PDF upload size

☁️ Deployment (Hugging Face Spaces)

You can easily host this application for free on Hugging Face Spaces using Docker.

⚠️ Important note on Deployment vs Local: The main branch of this repository contains the full application, including CLIP and EasyOCR for image understanding. However, these models require ~2.7GB of RAM, which exceeds the free tier limits of most platforms (including Hugging Face Spaces' 16GB limit due to boot-time memory spiking and persistent disk limits).

To solve this, a huggingface/ directory is provided. It contains a "slimmed down" version of the app that removes the Image OCR/CLIP models. The text-based RAG chat works perfectly and fits within a ~300MB footprint.

How to deploy to Hugging Face Free Tier:

  1. Create a new Space on Hugging Face.
  2. Choose Docker as the Space SDK.
  3. Choose Blank template.
  4. Upload the contents of the huggingface/ directory (the modified rag.py, app.py, requirements.txt, and Dockerfile).
  5. Copy the templates/ and static/ directories from the root folder into the Space as well.
  6. The Space will automatically build and start the Docker container.
  7. Users can visit your Space, enter their own Google API key in the UI, and chat with PDFs for free.

🔒 Security Notes

  • All AI output is sanitised with DOMPurify before rendering to prevent XSS
  • API keys are loaded from .env and never exposed to the frontend
  • Uploaded filenames are sanitised and timestamped (uploaded_<timestamp>.pdf)
  • PDF caching uses MD5 content hashing — not filenames

📄 License

MIT License — see LICENSE for details.

Built with ❤️ using Flask, FAISS, and Google Gemini

About

A RAG QnA Project , Which assist users in reading long form documents with ease. Users can read imp topics by querying and directly ask for explanations or their doubts regarding whatever, meanwhile also having access to see the document, They also can save particular responses to their notes.

huggingface.co/spaces/Avish006/DocuMind

Resources

Readme
Activity

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

Packages

Used by

Contributors

Languages