Errors

Every error code this API can answer with, the HTTP status that carries it, and what to do about it. The envelope is OpenAI-shaped with one namespaced addition.

The envelope

error.message is written for a person reading a log, error.type names the class of failure, and error.code is the stable identifier to branch on — message text is not a contract. Where the API has structured detail (which policy, which endpoint, which model) it arrives in error.llmeu and never replaces the three standard fields.

{
  "error": {
    "message": "No endpoint satisfies residency eu-owned. Residency is never relaxed silently.",
    "type": "conflict_error",
    "code": "no_endpoint_for_policy",
    "llmeu": { "requested": "llmeu-auto" }
  }
}

type follows status

The mapping below is the one the API applies, rendered from the same table that builds its responses.

HTTP	type
400	invalid_request_error
401	authentication_error
402	insufficient_quota
403	permission_error
404	not_found_error
409	conflict_error
413	invalid_request_error
422	invalid_request_error
429	rate_limit_error
500	api_error
501	api_error
502	api_error
503	api_error

Codes

Every code that can reach a client. A test walks the API source and fails when this table and the code disagree in either direction, so a code cannot be invented here or added there without being documented.

HTTP	code	What it means, and what to do
400	invalid_request	The request body failed validation.
400	invalid_json	The body was not valid JSON.
400	invalid_llmeu	The llmeu object was not an object, or carried an unknown field.
400	invalid_residency	The requested residency is not one of the four hosting classes.
400	invalid_data_class	The requested data_class is not one of the four classes.
400	invalid_retention	The requested retention is not zero, 24h, 7d or 30d.
400	invalid_max_tokens	max_tokens was not a positive integer.
400	invalid_stream	stream was not a boolean.
400	invalid_stop	stop was neither a string nor an array of strings.
400	invalid_role	A message carried a role outside system, user, assistant and tool.
400	invalid_tools	tools or tool_choice was malformed.
400	missing_model	The model field is required.
400	missing_messages	messages is required and must not be empty.
400	missing_input	input is required for embeddings.
400	missing_name	name is required.
401	invalid_api_key	The key is missing, unknown, revoked or expired.
402	budget_exceeded	The organization's daily budget is spent. Raise max_usd_per_day or wait for the window to roll.
403	api_key_required	The request used a console session. Create an API key and send that instead.
403	insufficient_scope	The key is valid but lacks the scope this endpoint requires.
403	no_organization	The credential is not attached to an organization.
403	model_denied	The organization's policy denies this model or its provider.
403	policy_violation	The request violates a rule of the organization policy.
403	retention_exceeds_policy	The requested retention is longer than the policy allows. This is checked before anything else about the request.
403	data_class_not_allowed	The request's data_class is not one the policy allows.
404	model_not_found	No such model for this organization.
404	hosted_unavailable	The model is in the catalogue but not hosted on LLM EU, so it cannot be called.
404	not_found	The route exists but the object does not.
404	unknown_route	No route is mounted at this path. Every OpenAI endpoint that is not listed under Endpoints answers this.
404	policy_not_found	The policy_id in llmeu does not belong to this organization.
409	no_endpoint_for_policy	Nothing satisfies the policy and the residency together, and neither is relaxed silently.
409	not_a_text_model	The named model cannot generate text: it is a speech-recognition, embedding or reranker model.
409	conflict	The request conflicts with the current state of the object.
413	request_too_large	The body exceeded the endpoint's size limit.
429	rate_limit_exceeded	A limit from the rate-limit table was reached. The message names the window reset time.
500	internal_error	An unexpected failure on our side.
501	not_implemented	The endpoint is intentionally not offered, or the requested capability does not exist on this deployment.
501	retention_not_implemented	The policy allows a retention other than zero and only zero is implemented, so the request is refused instead of being stored.
501	unsupported_n	n greater than 1 is not supported.
502	upstream_error	No endpoint could serve the request after failover. The trace records every attempt.
502	stream_failed	The stream failed after it had already started.

no_endpoint_for_policy is the product working

A 409 with this code means the policy and the request together admit no endpoint, and the router refused to relax either one silently. It is not transient: retrying without changing the request, the policy or the residency fails the same way. The error carries the requested residency, and the trace records why each endpoint was rejected.

Errors inside a stream

A streaming request that fails after the headers are sent writes the error as a final SSE frame and then the terminator, so a client that follows the format sees a failure instead of a truncated success.

data: {"error":{"message":"No endpoint could start the stream.","type":"api_error","code":"upstream_error"}}

data: [DONE]