Chat completions API: OpenAI-compatible and processed in the EU

POST /v1/chat/completions takes a conversation and returns the next assistant message. The conversation is a list of messages, each with a role (system, user, assistant or tool) and content. The model reads the whole list on every call, so the API itself is stateless: to continue a conversation, append the reply and the next user message and send the list again.

Because the format matches OpenAI's, the official Python and JavaScript SDKs, LangChain, LlamaIndex and most agent frameworks can call it without an adapter. You choose the model with the model field, so switching between open models is a one-string change.

Only model and messages are required. max_tokens caps the length of the reply, and temperature and top_p control how varied it is: keep temperature low for extraction and classification, higher for drafting. stop ends generation at a given string, and seed makes sampling more repeatable when you need stable test output.

Set stream to true to receive tokens as they are generated. Add tools to let the model call your functions, and response_format to force JSON that matches a schema. Images go into content as image_url parts when you use a vision model. Parameters a model does not support are ignored or rejected with a clear error.

A completed request returns an object with an id, the model that answered, and a choices array. The reply text is in choices[0].message.content. If the model decided to call a tool, message.tool_calls holds the function name and JSON arguments instead.

finish_reason tells you why generation stopped: stop means the model finished naturally, length means it hit max_tokens, and tool_calls means it is waiting for your tool results. The usage object reports prompt_tokens, completion_tokens and total_tokens, which is what you are billed on and what you should log to watch costs per feature.

Most product features follow the same shape: a fixed system message that sets tone and rules, a few lines of retrieved context, and the user's question. Keep the system message stable across requests so that prompt caching can reuse it, and put variable content at the end.

For anything a program reads, such as tags, routing decisions or extracted fields, combine a low temperature with response_format so you get valid JSON rather than parsing prose. For chat interfaces, stream the reply so users see text straight away, and trim old turns or summarise them once the conversation approaches the model's context window.

Errors use the OpenAI shape: an HTTP status and an error object with message, type and code. A 400 usually means an invalid field or a prompt longer than the model's context window. A 401 means the key is missing or wrong, 404 an unknown model id, and 429 that you have reached a rate limit for your plan.

Check finish_reason on every response. A length result means the answer was cut off, so raise max_tokens or ask for a shorter answer. Retry 429 and 5xx responses with exponential backoff, and honour the Retry-After header when it is present.