Choose a model provider

Every extractor selects a provider and model. Start with the path that matches where you want inference to run:

Path	Best when	Setup
Bundled vLLM runtime	You want ParseHawk’s default, fully local NuExtract3 setup	Use the bundled runtime
Ollama	You want a simple local model manager or a small vision-language model	Use Ollama locally
OpenAI	You want a first-party OpenAI model	Use OpenAI
Microsoft Foundry	Your organization deploys models through Foundry	Use Microsoft Foundry
Any compatible server	You operate LM Studio, a separate vLLM server, or another implementation	Use an OpenAI-compatible API

The bundled runtime, Ollama, and generic servers share the openai_compatible_api provider slot. Changing that slot changes the endpoint for every extractor assigned to it. Use separate ParseHawk installations when you must target multiple compatible endpoints at the same time.

Inspect current configuration

parsehawk providers list
parsehawk providers get openai_compatible_api
parsehawk providers models openai_compatible_api

API keys are encrypted at rest and are never returned by the API.

Skip the bundled runtime

If every active extractor uses a separate endpoint, avoid loading NuExtract3:

parsehawk start -x runtime

Use this for every provider path except the bundled runtime. The API, worker, Web UI, and Phoenix still start normally.

Assign a provider per extractor

Provider configuration stores connection details. The extractor stores which slot and model it should use:

parsehawk extractors update invoice_v1 \
  --provider openai_compatible_api \
  --model my-model

Existing jobs retain their recorded execution metadata. New jobs use the extractor’s current provider and model.