Use Microsoft Foundry

The microsoft_foundry provider separates the inference endpoint from the project endpoint used to discover deployments.

Collect the two URLs

You need:

• the OpenAI-compatible endpoint ending in /openai/v1
• the project endpoint ending in /api/projects/<project-name>
• an API key allowed to call the deployment and read project deployments

Configure the provider

export MICROSOFT_FOUNDRY_API_KEY=...

parsehawk providers configure microsoft_foundry \
  --base-url https://resource.services.ai.azure.com/openai/v1 \
  --project-url https://resource.services.ai.azure.com/api/projects/project-name \
  --api-key-env MICROSOFT_FOUNDRY_API_KEY

Inspect the stored non-secret configuration and list compatible deployments:

parsehawk providers get microsoft_foundry
parsehawk providers models microsoft_foundry

Assign the deployment

The extractor’s model is the chat-completions deployment name, not necessarily the underlying catalog model name:

parsehawk extractors update invoice_v1 \
  --provider microsoft_foundry \
  --model my-chat-deployment

Choose a deployment that supports structured chat completions. Image and PDF inputs also require image capability.

Start without the bundled runtime

parsehawk start -x runtime

Provider keys remain encrypted in the shared ParseHawk data directory. Both API and worker must use the same database and secret-key source.