Raw prompt mode

Send a custom system/user prompt straight to the model instead of a named pipeline — request shape, model selection, and the constraints that don't apply to named pipelines.

Every other guide in this section calls a named pipeline (translate-string, summarize, etc.) — a fixed prompt template with declared inputs. Raw mode skips the template: you send your own system_prompt and user_prompt directly, and the platform runs it as-is. Use it for prototyping a prompt before it’s worth turning into a real pipeline, or for a one-off call that doesn’t fit any of the built-in tools.

Getting it enabled

Raw mode is gated separately from named pipelines. Your JWT’s pipelines[] claim needs an explicit "raw" entry — the same claim used for named pipeline IDs, just with this one extra string. It is not granted by default, and most end-user session tokens should never have it: a named pipeline’s cost is bounded by its template, but a raw prompt’s cost is bounded only by how long a prompt you’re willing to send and how many tokens you ask the model to generate.

If your app mints session tokens itself (see Node.js server integration’s POST /auth/token section), request "raw" only for the specific tokens that need it — e.g. an internal admin tool, not tokens handed to arbitrary end users. Ask your platform operator to confirm your Application’s allowed_pipelines includes "raw" before you request a token with that claim; a claim your Application isn’t allowed to grant is silently dropped from the issued token.

A static x-api-key (unbound, operator-scoped — see Get your API key) bypasses the pipeline-list check entirely, same as it does for named pipelines. A Console-issued, app-bound x-api-key and a JWT/webtoken caller both need the explicit "raw" grant.

Calling it — SDK

const out = await ai.run({
  raw: {
    system_prompt: "You are a terse assistant. Answer in one sentence.",
    user_prompt: "What does raw mode skip that a named pipeline has?",
    model: "gpt-5-mini",   // optional — see "Choosing a model" below
    temperature: 0.2,       // optional
    max_tokens: 500,        // optional, up to 32768
  },
});
console.log(out.text);

user_prompt can also be a function of inputs when used with ai.runMany(...) — see the SDK’s fuller Integration Guide (ask your platform operator for a copy — it isn’t public) for that pattern. ai.button(...) also accepts a raw config in place of pipeline.

Calling it — direct HTTP

Requires Authorization: Bearer <jwt> with the "raw" grant (a static x-api-key works too if it’s the unbound operator kind — see above). params is optional and both fields inside it are optional:

curl -X POST "https://api.quravin.com/tickets" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <your-jwt>" \
  -d '{
    "mode": "raw",
    "system_prompt": "You are a terse assistant. Answer in one sentence.",
    "user_prompt": "What does raw mode skip that a named pipeline has?",
    "model": "gpt-5-mini",
    "params": { "temperature": 0.2, "max_tokens": 500 }
  }'

Response (202), then poll the same way as any ticket — see Direct API integration if you haven’t built that loop yet:

{ "ticket_id": "01J...ULID", "status": "QUEUED" }

curl "https://api.quravin.com/tickets/01J...ULID" \
  -H "Authorization: Bearer <your-jwt>"

Response once done — the shape is always the same four fields plus text, regardless of which model ran it:

{
  "ticket_id": "01J...ULID",
  "status": "DONE",
  "mode": "raw",
  "result": {
    "text": "It skips the fixed prompt template and declared input schema.",
    "model": "gpt-5-mini",
    "latency_ms": 640,
    "tokens_in": 38,
    "tokens_out": 14
  },
  "created_at": "2026-07-10T12:00:00.000Z",
  "updated_at": "2026-07-10T12:00:02.000Z"
}

There’s no pipeline-specific field to look up (no translation, no tldr) — result.text is always where the model’s output lands.

Choosing a model

model is optional — omit it to use the platform’s current default. If you set it explicitly, it must be one the platform has a verified cost rate for, checked before the call reaches the LLM: send anything else and you get 400 invalid_model instead of an expensive surprise. As of this writing that list is gpt-4o-mini, gpt-4o, o3, and gpt-5-mini — ask your platform operator for the current list, since it changes as models are added.

What’s different from a named pipeline

No input schema. A named pipeline validates your inputs against its declared schema before the call is accepted (missing field, wrong type → 400). Raw mode has no schema — system_prompt and user_prompt are required strings, but their content is never validated. A prompt that doesn’t make sense to the model doesn’t fail upfront; the ticket runs, and you find out when you poll the result.
No caching. Named pipelines cache identical calls by their inputs. Raw mode never checks a cache — every call is a fresh model call, every time, even if you send the exact same prompt twice in a row.
The global input-size backstop still applies, measured differently: system_prompt.length + user_prompt.length combined must stay under the platform’s character cap (50,000 by default) — same defense-in-depth limit that bounds named-pipeline inputs, just added up across the two fields instead of the whole JSON body.
max_tokens is capped at 32768 regardless of what the underlying model would otherwise allow — a ceiling on a single call’s worst case, not a target to aim for.

Errors

Same authenticated-path error table as any other call — see Direct API integration for the full list. One error is specific to raw mode: 400 invalid_model, covered above.

If your JWT doesn’t carry the "raw" grant, you get the same 403 as calling a named pipeline you’re not allowed to use — see Choose your auth mode for how credential grants work.