Response language

Bayle renders error messages in English or Brazilian Portuguese. Successful responses carry generated content, which is unaffected by this: it is only error prose that gets translated.

Asking for a language

Send Accept-Language. A tag beginning with the two letters pt, in any case, selects Brazilian Portuguese. Anything else, including no header at all, selects English.

curl https://api.bayle.ai/embeddings \
  -H "Authorization: Bearer bad-key" \
  -H "Accept-Language: pt-BR" \
  -H "Content-Type: application/json" \
  -d '{"model": "bad-model-id", "input": "hello"}'

{
  "error": {
    "message": "Sua chave de API da Bayle está ausente ou não pôde ser verificada.",
    "type": "authentication_error",
    "param": null,
    "code": "invalid_api_key"
  }
}

The tag may appear anywhere in a multi-tag header, so fr-FR;q=0.9,pt;q=0.5,en;q=0.3 selects Portuguese. Quality values are accepted but ignored: this is a presence check for a pt-prefixed tag, not weighted negotiation. pt, pt-BR, and pt-PT all select the same Brazilian Portuguese messages.

Knowing which language you got

Every response carries Content-Language, either en or pt-BR, on success and error alike. You can read it without parsing message text.

curl -sD - https://api.bayle.ai/models -o /dev/null | grep -i content-language

On a streaming chat completion the header is committed alongside the initial 200, before the outcome is known, so it is present whichever way the stream ends. An error event arriving mid-stream is rendered in the language negotiated at the start of the request.

Don’t match on message text

Because message is translated, code that matches on it works until someone changes a header. Branch on error.code, which is stable across both languages. See Errors.