Set up an OpenAI-compatible model

Many services and servers implement the OpenAI chat completions API, so one provider covers them all: you only choose the service and the model. Compass asks the model for probabilities as JSON and validates them.

Caution

Chat models state probabilities they have not calibrated, and they can change between model versions. Use this provider for reviews in the dev toolbar, and prefer TypeSafe when you fail builds on findings.

Prerequisites

• Compass installed in your Starlight site.
• An API key for a service that implements the API, or a local server like Ollama running a model with support for structured outputs.

Configure the provider

1. Add the plugin with the openaiCompatible() provider to your Starlight configuration. Choose the service:

   OpenAI

   astro.config.mjs

   import starlightCompass, { openaiCompatible } from "starlight-compass";
   
   starlightCompass({
     provider: openaiCompatible({ model: "gpt-4o-mini" }),
   });

   OpenRouter

   astro.config.mjs

   import starlightCompass, { openaiCompatible } from "starlight-compass";
   
   starlightCompass({
     provider: openaiCompatible({
       baseUrl: "https://openrouter.ai/api/v1",
       model: "deepseek/deepseek-chat",
     }),
   });

   Ollama

   astro.config.mjs

   import starlightCompass, { openaiCompatible } from "starlight-compass";
   
   starlightCompass({
     provider: openaiCompatible({
       baseUrl: "http://localhost:11434/v1",
       model: "llama3.1",
     }),
   });

2. Create a .env file in your project root with your API key, and make sure .env is listed in your .gitignore file. Local servers like Ollama need no key, so skip this step for them.

   .env

   OPENAI_API_KEY=sk-...

   The provider reads the OPENAI_API_KEY variable whatever the service is. Pass the apiKey option to read another variable.

Next, review your first page.