Skip to content

Local development

Work with ML models

When uploading a proof, machine learning models are run asynchronously in the background, thanks to Django Q2.

However, it still requires a Triton server to be running, as well as a Gemini API key to be set in the environment variables.

Triton is used to:

  • detect the proof type (proof_classification)
  • detect price tags on the proof (price_tag_detection)

Gemini is used to:

  • extract information from receipts (receipt_extraction)
  • extract information from price tags (price_tag_extraction)

To run Triton, you should first clone the Robotoff repository.

Then, download the object detection models: 

make dl-object-detection-models

Then, run the Triton server:

make up service=triton

You can check with docker logs triton that the server is running correctly.

For Gemini, the easiest way to set it up is to use the Gemini API. You can set your API key by overriding the GEMINI_API_KEY environment variable in your .envrc file:

export GEMINI_API_KEY=your_api_key_here