🧠 AskScribe — AI-Powered Document Intelligence System

AskScribe is an intelligent document analysis tool built with Flask, combining Google Gemini AI and FAISS vector search to help users ask questions about uploaded documents in natural language. Whether you're scanning PDFs, DOCX, or TXT files — AskScribe extracts, indexes, and intelligently answers with structured, professional responses.

---

🚀 Features

• 📂 Upload PDFs, DOCX, and TXT files
• 🧠 Ask questions and get structured, context-aware answers
• 🧾 Supports multi-session chat history
• 🔍 Custom TF-IDF + FAISS vector search engine
• 🖼️ OCR support for scanned documents
• 💬 Gemini-powered LLM responses with Markdown formatting
• 🔐 Secure user authentication and session handling
• 📁 Embedded file management, chunking, and vector indexing

---

📁 Project Structure

DocumentIntelligence/
│
├── templates/              # HTML (Jinja2)
├── static/                 # CSS/JS/Assets
├── uploads/                # Uploaded documents
├── vectors/                # Stored vector index (JSON)
│
├── main.py                 # Entry point
├── routes.py               # App routes
├── gemini_client.py        # Gemini integration
├── rag_engine.py           # Vector search & RAG engine
├── models.py               # SQLAlchemy models
├── utils/                  # OCR, chunking, preprocessing
├── requirements.txt        # Dependencies
└── README.md

---

⚙️ Getting Started

1️⃣ Clone & Setup

git clone https://github.com/yourusername/askscribe.git
cd askscribe
python -m venv venv
source venv/bin/activate  # or venv\Scripts\activate on Windows
pip install -r requirements.txt

2️⃣ Environment Variables

Create a .env file and add:

GEMINI_API_KEY=your_google_gemini_key
SESSION_SECRET=your_flask_secret

3️⃣ Run the App

python main.py

Visit http://localhost:5000 in your browser.

---

🧠 How It Works

📂 Document Upload

Users upload files → Extract content → Chunk → Generate TF-IDF embeddings → Store with FAISS

❓ Question Answering

User asks question → Retrieve top relevant chunks → Construct prompt → Gemini generates answer

💬 Chat Interface

Real-time Q&A → History stored per session → View or continue previous chats

---

🧰 System Architecture

🎨 Frontend Architecture

• Framework: Bootstrap 5.3.2 with dark theme
• Styling: Custom CSS with ChatGPT-inspired dark interface
• JavaScript: Vanilla JS with Bootstrap components
• Templates: Jinja2 templating engine
• Features: Responsive design, file upload validation, real-time chat interface

🧰 Backend Architecture

• Framework: Flask with SQLAlchemy ORM
• Authentication: Flask-Login for session management
• Database: SQLite (configurable via DATABASE_URL)
• File Processing: Multi-format document processing with OCR
• AI Integration: Google Gemini 2.5 Flash model
• Vector Search: FAISS for efficient similarity search

---

🔍 Key Components

📄 Document Processing Pipeline

• Text Extraction: Full support for PDF (PyMuPDF), DOCX, and TXT
• OCR: Automatic Tesseract fallback for scanned documents
• Chunking: 1000-char chunks with 200-char overlap
• Embeddings: Custom TF-IDF embeddings (lightweight)

⚙️ RAG Engine

• Vector Store: JSON-based TF-IDF + FAISS similarity
• Retrieval: Cosine similarity for top-matching chunks
• LLM Generation: Prompt engineering with Gemini AI

🔐 Authentication

• User System: Registration, login, logout
• Session Handling: Secure cookie-based sessions with Flask-Login
• Protected Routes: Auth-required pages

💬 Chat System

• Session Support: Multi-chat session per user
• Chat History: Stored persistently in database
• Contextual Q&A: Gemini answers are always based on document context

---

🔁 Data Flow

flowchart TD
  A[User Uploads Document] --> B[Text Extraction + OCR]
  B --> C[Chunking & Embedding]
  C --> D[FAISS Vector Storage]
  D --> E[Ask Question]
  E --> F[Relevant Chunks Retrieved]
  F --> G[Prompt Sent to Gemini AI]
  G --> H[Response Returned to User]

Loading

---

---

🛠️ Tech Stack

Layer	Tools / Libraries
🧠 AI Model	Google Gemini 2.5 Flash
🔍 Search	FAISS + TF-IDF (custom implementation)
🧾 OCR	Tesseract + Pillow
🧰 Backend	Flask, SQLAlchemy, SQLite
🎨 Frontend	HTML, Bootstrap 5.3, JS
🔐 Auth	Flask-Login

---

🧪 Sample Gemini Prompt

**Question**: What is the policy on leave?
**Context**: [Relevant chunks retrieved]
**Instructions**: Answer with headings, bullet points, and highlight **key terms**.

---

🔐 Security Features

• ✅ Secure file storage with size/type checks
• ✅ CSRF protection & secure sessions
• ✅ Environment-based secrets (no hardcoding)
• ✅ Auto OCR fallback for scanned documents

---

🔮 Future Upgrades

📤 Chat Export

Export session as PDF, Markdown, or TXT for offline sharing.

☁️ Cloud Uploads

Switch to Amazon S3 or Google Cloud Storage for large files.

📈 Analytics Dashboard

Track document types, most asked queries, usage trends.

🔔 Real-Time Notifications

Get alerts for OCR status, Gemini API limits, and timeouts.

---

📄 License

This project is licensed under the MIT License. See the LICENSE file for more info.

---

👨‍💻 Author

Made with ❤️ by Md Meraj Alam
Your feedback is always welcome!