Choose where a request runs

Your workspace policy sets the rule for every request. Sometimes one request needs a tighter rule. An example: your own customer in Germany needs EU compute, and a customer elsewhere does not. The provider object sets that rule for one request.

The three fields

region is another name for runs_in, with the same meaning. It is not a provider region code. To pick one of a provider's regions, put it in the model id: creator/model/provider:<code>.

Use the slugs that GET /v1/models/{model}/endpoints lists for the model.

What it guarantees

• It only narrows. The request runs inside your workspace policy and inside the filter. It can never reach an endpoint that your workspace policy excludes.
• It applies to the endpoint, not the model. only: ["ovhcloud-ai-endpoints"] uses OVHcloud's endpoint, whichever company is listed as the model's provider.
• Short names work too. Some providers also take a short name, for example ovh for ovhcloud-ai-endpoints. It is the same short name a pin takes (baai/bge-m3/ovh), and the request log shows the full slug.
• Failover stays inside it. If the first endpoint fails, Stav tries the model's next endpoint that passes the filter. If no endpoint passes, you get Stav's own error. The request is never sent outside the filter to rescue it.
• Nothing is sent when nothing passes. You get a refusal and the request does not reach any provider.

Chat

from openai import OpenAI

client = OpenAI(base_url="https://api.stav.ai/v1", api_key="sk_stav_live_...")

resp = client.chat.completions.create(
    model="mistralai/mistral-small-4-119b-2603",
    messages=[{"role": "user", "content": "Summarise this contract clause."}],
    extra_body={"provider": {"runs_in": "eu"}},
)

This works the same with model="auto" and with a Router. The Smart Router only scores endpoints inside the filter.

Messages (Anthropic SDK)

import anthropic

client = anthropic.Anthropic(base_url="https://api.stav.ai", api_key="sk_stav_live_...")

msg = client.messages.create(
    model="mistralai/mistral-small-4-119b-2603",
    max_tokens=512,
    messages=[{"role": "user", "content": "Hei!"}],
    extra_body={"provider": {"runs_in": "NO"}},
)

Most Claude endpoints are global. global is not in any region, so runs_in: "eu" refuses them until an EU endpoint for that model exists.

Responses

resp = client.responses.create(
    model="openai/gpt-5",
    input="Hello",
    extra_body={"provider": {"ignore": ["azure"]}},
)

Stav takes provider off the body. It is never sent to the provider.

Embeddings

emb = client.embeddings.create(
    model="baai/bge-m3",
    input=["Hovedstaden i Norge er Oslo."],
    extra_body={"provider": {"runs_in": "eu"}},
)

Rerank

curl https://api.stav.ai/v1/rerank \
  -H "Authorization: Bearer $STAV_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "baai/bge-reranker-v2-m3", "query": "Where is Stavanger?",
       "documents": ["Stavanger is in Norway.", "Paris is in France."],
       "provider": {"runs_in": "eu"}}'

Speech and decisions

/v1/audio/speech and /v1/decisions take the same provider object in the JSON body.

Transcriptions

/v1/audio/transcriptions uses a multipart form. Send provider as one form field that holds a JSON object:

curl https://api.stav.ai/v1/audio/transcriptions \
  -H "Authorization: Bearer $STAV_API_KEY" \
  -F file=@meeting.wav \
  -F model=mistralai/voxtral-mini-2602 \
  -F 'provider={"runs_in": "eu"}'

Fields that apply on chat only

max_price and quantizations filter the endpoints of a chat request, like the fields above: max_price is a price per million tokens, and embeddings, rerank, audio and decisions are not all billed that way. sort orders a named model's endpoints on the chat surfaces only (see Routing hints). Stav refuses all three there with 400 unsupported_parameter.

When the request is refused

Check it in the response

Every response says where it ran, so your code can check a runs_in: "eu" call with no second request.

global means the Inference Provider pins no residency. unknown means Stav holds no fact about where the endpoint runs. They are two different values.

A call ran in the EU when the region is eu, or when every country is an EU state. Speech answers with audio, so it has the headers only. A transcription in text, srt or vtt has the headers only too.

See what happened

Every request with a filter shows it in Your workspace → Monitor → Request log. Open the request to see the filter, the endpoint that served, and the country it ran in.

Next steps

• Enforce a sovereignty floor — the workspace rule this narrows
• Sovereignty
• Routing hints