⚡ Typhoon OCR API (Optimized)

High-performance Optical Character Recognition (OCR) API built on top of the SCB10X Typhoon OCR 7B model, powered by FastAPI and Transformers.

Designed for scalable deployment and optimized for GPU acceleration and low-latency inference.

---

🚀 Features

• 🔍 General and custom prompt-based OCR
• 🧠 Powered by Typhoon OCR 7B (Multimodal vision-language model)
• ⚡ GPU optimization: TF32, cuDNN tuning, memory handling
• 📷 Efficient image resizing and format conversion
• 📤 Upload image via REST API with prompt or message support
• 📦 Fully containerizable for deployment

---

📦 Installation

1. Clone Repository

git clone https://github.com/your-org/typhoon-ocr-api.git
cd typhoon-ocr-api

2. Install Dependencies

Make sure Python 3.10+ is installed.

pip install -r requirements.txt

✅ For GPU support, ensure that you have CUDA and the proper torch version installed. Example:

pip install torch==2.1.0+cu118 -f https://download.pytorch.org/whl/torch_stable.html

---

🧪 Usage

Run the server

uvicorn main:app --host 0.0.0.0 --port 8000

Then open your browser at: 📍 http://localhost:8000/docs

---

📤 API Endpoints

POST /ocr/general

OCR with a general prompt (default or custom string)

Request:

• file: image file (jpg, png, etc.)
• prompt: optional, custom instruction (default: "Extract all text from this image")

Example:

curl -X POST http://localhost:8000/ocr/general \
  -F "file=@./sample.jpg" \
  -F "prompt=Extract the ID number and name from the ID card"

---

POST /ocr/custom

OCR with a structured prompt using role-based messages.

Request:

• file: image file
• messages: JSON list of chat-style messages

Example Body (in Swagger or Postman):

[
  {
    "role": "user",
    "content": "Extract text in table format from the document"
  }
]

---

GET /healthz

Simple health check endpoint. Returns "ok" if the model and processor are loaded correctly.

---

⚙️ Environment Requirements

• Python >= 3.10
• CUDA-enabled GPU (optional but recommended)
• Pytorch >= 2.1.0
• Transformers >= 4.40.0

---

📊 Logging

Application logs are printed to stdout using Python logging. Includes:

• Model load time
• GPU availability and memory
• OCR process time

---

📌 Notes

• This API loads the Typhoon OCR 7B model and performs warm-up during startup.
• GPU optimization includes TF32 acceleration, cuDNN tuning, and torch.compile for inference.
• You can extend the API to support multiple models, batch processing, or file exports easily.

---

🧑‍💻 Author

Sitthichai S. (2025)

---

📄 License

MIT License © 2025

---