OpenAI compatibility

Compatibility is a floor, not a feature. It is here so that moving to Grid costs you two lines rather than a rewrite, and so that leaving would cost you the same. What Grid asks you to evaluate is portability, routing and the meter, not the shape of the request body.

Using the OpenAI SDKs#

Set the base URL and use the client you already have. Authentication is a bearer token in the standard Authorization header, so nothing about the auth flow changes.

import os
import openai

client = openai.OpenAI(
    base_url="https://api.genai.forlex.cloud/v1",
    api_key=os.environ["FORLEX_API_KEY"],
)

import OpenAI from 'openai'

const client = new OpenAI({
  baseURL: 'https://api.genai.forlex.cloud/v1',
  apiKey: process.env.FORLEX_API_KEY,
})

If you prefer environment variables, OPENAI_BASE_URL and OPENAI_API_KEY work as they normally do, which lets you point an existing application at Grid without touching its code at all.

Endpoints#

GET /v1/models returns the OpenAI list envelope, so SDK helpers that enumerate models work unchanged:

{
  "object": "list",
  "data": [
    { "id": "livia-e1-flash", "object": "model", "created": 1700000000, "owned_by": "Forlex" },
    { "id": "livia-e1-mini", "object": "model", "created": 1700000000, "owned_by": "Forlex" }
  ]
}

The list is the authority on what you can call. Prefer reading it at runtime over hardcoding a model id, both because the catalog is curated and because it is the one source that cannot be out of date.

Parameters on chat completions#

Parameters not listed here are not documented as working. That is a statement about what we have verified, not a statement that they fail. If you need one confirmed before you build on it, ask and we will answer specifically rather than generally.

Where Grid differs#

Two differences are worth knowing before you write error handling.

Authentication failures return a plain text body. A request with a missing or invalid key answers 401 with content-type: text/plain, not the JSON {"error": {...}} envelope the OpenAI SDKs expect. The SDKs will surface the status code correctly but may not parse it into a typed error object. If your code branches on a parsed error body, handle the status code as well.

Serverless can answer 503 on a cold lane. The OpenAI API does not scale to zero, so this is a state your existing retry logic may not cover. See Errors and cold start.

Response headers#

Grid adds x-grid-routed-model, naming the model that produced the response. It is present whether or not you used routing, which makes it the single thing to log if you want to be able to reconstruct an answer later.