Web API & demo app

The repository contains two ready-to-run applications: a REST API template to integrate docTR into your services, and an interactive demo to try the models.

REST API

The api/ folder is a FastAPI application exposing docTR over HTTP, packaged with Docker.

Start the server

You need Docker and Python (to resolve the dependencies with uv):

git clone https://github.com/mindee/doctr.git
cd doctr/api
make run

The server listens on port 8080. Its interactive documentation (Swagger UI) is available at http://localhost:8080/docs.

Routes

Route	Description
POST /detection	Text detection: the boxes of the words
POST /recognition	Text recognition of word crops
POST /ocr	End-to-end OCR
POST /kie	Key information extraction

Each route accepts images and PDFs (application/pdf, image/jpeg, image/png) as multipart files, and the architectures to use as query parameters:

import requests

params = {"det_arch": "fast_base", "reco_arch": "crnn_vgg16_bn"}

with open("path/to/invoice.pdf", "rb") as f:
    files = [("files", ("invoice.pdf", f.read(), "application/pdf"))]

response = requests.post("http://localhost:8080/ocr", params=params, files=files)
print(response.json())

The template is a starting point: adapt the routes, the response schemas and the models to your needs. The API README documents every route with examples.

Demo app

A Streamlit app lets you upload a document, pick the models, and inspect the segmentation maps, the detected words and the reconstructed page.

Try it online on Hugging Face Spaces, or run it locally from the repository:

pip install -r demo/pt-requirements.txt
streamlit run demo/app.py

Example script

scripts/analyze.py analyzes a single file and displays the results, a handy way to try options from the terminal:

python scripts/analyze.py path/to/your/doc.pdf
python scripts/analyze.py --help

For batch processing from the terminal, use the CLI.