AI Document Enhancement & Intelligent Chat System

An end-to-end Document Intelligence Platform that enhances, extracts, and enables intelligent interaction with documents using Deep Learning and Large Language Models.

---

Overview

This system allows users to:

• 1 Upload PDF or image documents
• 2 Enhance document quality (Deblur + Super Resolution)
• 3 Extract text using AI-powered Vision OCR
• 4 Store metadata and extracted text in MySQL
• 5 Select single or multiple documents
• 6 Chat with documents using real-time streaming LLM
• 7 Track document processing status

The platform combines Computer Vision, Deep Learning, OCR, and LLM-based reasoning into a unified, scalable architecture.

---

System Architecture

User Upload
    ↓
Store File Locally
    ↓
Save Metadata in MySQL
    ↓
Preprocess (DE-GAN + RealESRGAN)
    ↓
Save Enhanced File
    ↓
OCR (OpenAI Vision)
    ↓
Save Extracted Text
    ↓
User Selects Documents
    ↓
Streaming LLM Chat (LLAMA 3.2(open source))

---

Tech Stack

Layer	Technology
Backend	FastAPI
Database	MySQL
Enhancement	DE-GAN + RealESRGAN
OCR	OpenAI GPT-4o Vision
Chat LLM	LLAMA 3.2
Streaming	FastAPI StreamingResponse
PDF Handling	PyMuPDF
Image Processing	ocr
Server	Uvicorn

---

Project Structure

DE-GAN/
│
├── main.py
├── db.py
├── ocr_helper.py
├── degain_esrgan_pipeline.py
├── enhance.py
│
├── routers/
│   ├── upload_api.py
│   ├── preprocess_api.py
│   ├── ocr_api.py
│   ├── chat_api.py
│   ├── file_retriver.py
│   └── list_documents.py
│
├── image_for_enhancement/
├── realesrgan_output/
├── temp_pages/
├── templates/
└── weights/

---

Database Schema

Table: common

Stores document metadata and processing results.

CREATE TABLE common (
    id INT AUTO_INCREMENT PRIMARY KEY,
    preproc_path VARCHAR(255) NOT NULL,
    extract_text TEXT NOT NULL,
    usrdoc_name VARCHAR(255) NOT NULL,
    doc_name VARCHAR(255) NOT NULL,
    size INT NOT NULL,
    doc_status VARCHAR(255) NOT NULL,
    file_path VARCHAR(255) NOT NULL,
    time_stamp VARCHAR(255) NOT NULL
);

Table: querries

Stores chat history per document.

CREATE TABLE querries (
    doc_id INT NOT NULL,
    q_id VARCHAR(255) NOT NULL PRIMARY KEY,
    upload_time DATETIME NOT NULL,
    ext_time DATETIME NOT NULL,
    querry VARCHAR(255) NOT NULL,
    response VARCHAR(255) NOT NULL,
    FOREIGN KEY (doc_id) REFERENCES common(id)
);

---

API Endpoints

Upload Document

POST /upload

• Saves file locally
• Inserts metadata in DB
• Sets status = uploaded

Preprocess Document

POST /preprocess/{doc_id}

• Convert PDF → Images
• Run DE-GAN (Deblur)
• Run RealESRGAN (Super Resolution)
• Rebuild multi-page PDF
• Update preproc_path
• Set status = preprocessed

OCR Extraction

POST /ocr/{doc_id}

• Load enhanced file
• Extract text using OpenAI Vision
• Save text in DB
• Set status = extracted

List Documents

GET /documents

Retrieve File

GET /retrieve/{doc_id}/file?file_type=output
GET /retrieve/{doc_id}/file?file_type=preproc

Chat with Documents

POST /chat

Request Body:

{
  "doc_ids": [1, 2],
  "question": "Summarize the key findings."
}

How It Works:

1. Fetch extracted text from selected documents
2. Combine into context
3. Send to GPT-4o
4. Stream response using StreamingResponse
5. Save query & response in DB

---

Real-Time Streaming

return StreamingResponse(generator(), media_type="text/event-stream")

Frontend receives incremental tokens in real time.

---

Features

• ✔ Multi-page PDF support
• ✔ Parallel document enhancement
• ✔ AI Vision OCR
• ✔ Real-time streaming LLM chat
• ✔ Multi-document reasoning
• ✔ Modular API architecture
• ✔ MySQL integration
• ✔ Status tracking

---

Installation

1. Install Dependencies

pip install fastapi uvicorn mysql-connector-python
pip install opencv-python torch torchvision
pip install pymupdf pillow
pip install openai

2. Run Server

uvicorn main:app --reload

Swagger UI available at: http://127.0.0.1:8000/docs

---

Requirements

• Python 3.12.12+
• MySQL running with doc_man schema
• RealESRGAN weights inside /weights
• OpenAI API key configured( I have used open source model name LLAMA 3.2)

---

Future Improvements

• Background processing queue
• User authentication & role-based access
• Cloud deployment

---

Author

Talib Hussain Ansari
AI Engineer | Document Intelligence Systems

---

🏁 Summary

This project integrates Computer Vision, Deep Learning Enhancement, AI OCR, and LLM Streaming Chat into a scalable Document Intelligence platform designed for intelligent automation and interactive document analysis.