Skip to main content

Model backends

LangExtract does not hardcode a single model. The model_id you pass to lx.extract selects a provider (the integration that talks to a model), and different providers run on different infrastructure. This page explains how that selection works conceptually; for step-by-step setup, see the how-to guides linked below.

model_id selects a provider by pattern​

A provider registers the model-id patterns it handles, and LangExtract routes each model_id to the provider whose pattern it matches:

  • Gemini (the default). Model ids beginning with gemini route to Google Gemini. The default model_id is gemini-3.5-flash, so an extraction with no model_id runs on Gemini.
  • OpenAI. Model ids such as gpt-4o route to OpenAI. The integration is an optional extra you install separately.
  • Ollama (local). Model ids such as gemma2:2b route to Ollama, which runs models on your own machine with no API key.

Because routing is pattern-based rather than a fixed list, a provider can claim a family of models without LangExtract enumerating each one.

When a model id doesn't match​

If a model_id matches no registered pattern, LangExtract can't infer the provider and raises an error. You then name the provider explicitly through a configuration object. For the exact patterns, the precedence rules, and the environment variables each provider reads, see the API reference §4.

Extending with plugins​

Providers are pluggable: a third-party package can register new patterns through an entry point, adding support for more models without changing LangExtract itself. For details, see the API reference §4.

See also​