Separator
Separator is the routing layer. It does not try to be every model. It decides, in a checked and fast step, which model should do the work — and when the work should be split.
Decision plane
4.2 ms
explanationreason-1
Chat
Chat is the single-hop case. The request enters the gateway, receives a type, and is checked against a contract before a model is chosen. You see the type, the lane, and the decision. A second model is not invited in to round out the paragraph.
The packet below is drawn in the browser. It does not call reason-1.
Why is the sky blue?
reason-1
Explanation
Route<Question, Explanation>
When the message is mixed
Ask why the sky is blue and also ask for a patch. The shape changes from Route to Fanout. reason-1 still owns the explanation. code-1 owns the diff. The original question is not rewritten to include the file change.
Why is the sky blue, and patch the docs
reason-1 keeps the question
docs · code-1
+ Why the sky is blue
01 Chat
separator / chat
YouWhy is the sky blue?
Separatorexplanation → reason-1 · 3.2 ms
reason-1Shorter blue wavelengths scatter more off air molecules than longer red ones.
02 Build
14 export async function handler(req) {
15+ await limit(req, { max: 100 });
16 const session = await getSession(req);
17 return schema.parse(session);
type patch · Route<ChangeRequest, Diff>
03 Bot
fan-out
Close the month and file the report.
04 Imagine
media → vision-1 · Route<Prompt, Media>
05 Voice
speech-1
Transcribe the call and book the follow-up.
speech-1 + ops-1 · fan-out
Every request receives a type before a model is selected.
The routing choice is designed to land in under 10 ms.
reason-1, code-1, write-1, search-1, vision-1, speech-1, or ops-1.
A request with several jobs becomes several small requests.
Type, model, contract, and latency are part of the response.
A lane that is not allowed cannot be chosen.
Route<In, Out> for one hop. Fanout<Request, Part[]> for a split.
A type that fits no allowed lane returns pending and calls nothing.
Point a client at api.veloxity.si/v1 and set the model to separator.
Pass code-1, or any lane id, when the caller already knows the type.
The decision and the model’s generation are different charges.
VeloxitySI puts the router in the stack you already run.
01
Tell us which models you already call, and which kinds of requests show up.
02
Separator only routes to lanes you permit. A mismatch is a refusal.
03
Read the type, the model, and the parts. Switch on the contract in your app.
One hop
Chat on Separator is the single-hop case. The request is typed, checked against Route<In, Out>, and handed to the lane that answers that type. You see the type, the model, and the milliseconds spent choosing. You do not see a second model invited in to round out the paragraph.
If the same message also asks for a diff, a booking, or a picture, the shape changes to Fanout<Request, Part[]> and the extra jobs leave. The original question is still its own part.
Explanation, patch, prose, lookup, media, speech, or workflow. The type is assigned before a model is selected.
One type that fits an allowed lane is a Route. Several types are a Fanout. Neither is pending.
reason-1, code-1, write-1, search-1, vision-1, speech-1, or ops-1. A lane outside the allow-list is not a candidate.
Type, contract, model, latency, and parts. Prompt text is omitted unless you enable that logging.
A decision
The block is the shape, not a capture from production. latency_ms is the decision. It is not the time the models spend generating. The API is in preview. The consoles on this site draw the same shape in the browser and do not call a model.
Callers that already speak the OpenAI-compatible base can point at api.veloxity.si/v1 and set the model to separator. Callers that want a lane pinned send that lane id instead, and a mismatch is refused.
{
"id": "rte_01",
"latency_ms": 4.6,
"type": "fan-out",
"contract": "Fanout<Request, Part[]>",
"model": null,
"parts": [
{ "type": "workflow", "model": "ops-1" },
{ "type": "prose", "model": "write-1" },
{ "type": "explanation", "model": "reason-1" }
]
}Lanes
01 Chat
A plain question is classified, checked against a contract, and handed to the model that answers that kind of question. You see the type, the model, and the milliseconds.
02 Build
A change request is typed as a patch and routed to code-1. If the same message also asks for an explanation, that part splits off and does not ride inside the diff.
03 Bot
Separator reads a long ask, breaks it into typed parts, and calls a different model for each part. The list is the result. Nothing is merged in secret.
04 Imagine
Media is its own type. vision-1 receives the prompt. If you also ask for a caption, write-1 gets that part alone.
05 Voice
Audio is typed as speech and routed to speech-1. A follow-up hiding in the same utterance is split into its own call.
Outside the router
The answer, the diff body, the still, and the transcript take as long as the lane takes. That time is not folded into the decision target.
Applications stop keeping their own table of which screen calls which vendor. They post to one route and read the contract.
Nothing on this page is a SOC report, a HIPAA assessment, or a penetration test. The security statement says the same thing in formal language.
Under 10 ms is the design target for the routing choice. It is not a number we are publishing from a production trace on this website.
Separator
Decision
explanation
reason-1
4.2 ms