Inference API reference

# Integrate with the Webanto AI API

You are integrating an app with Webanto's inference API (chat + embeddings).
It is OpenAI-shaped but NOT the OpenAI API — follow this exact contract.

## Base URL and auth
- Base URL: https://api.webanto.com/api/v1
- Get an API key at https://webanto.com/account/settings/api-keys (format: wba_live_...).
  Keys are scoped per endpoint: "chat" and/or "embeddings".
- Send it as header: Authorization: Bearer wba_live_...  (or x-api-key: wba_live_...).
- Server-side only. Never expose the key in a browser or client bundle.

## POST /api/v1/chat  (scope: chat)
Request body:
    {
      "model": "qwen",
      "messages": [{ "role": "system|user|assistant", "content": "..." }],
      "max_tokens": 1024,
      "temperature": 0.7,
      "use_memory": false,
      "reasoning_effort": "none"
    }
- model must be "qwen" (multimodal: text + vision). messages: 1-50 turns.
- qwen is a REASONING model: keep max_tokens >= ~1024 or the visible output can be empty.
- reasoning_effort is OPTIONAL and defaults to reasoning ON.
  Set "none" for extraction, classification, or structured-output work: it skips the
  hidden reasoning phase, which cuts completion tokens (and therefore credits) by a
  large factor and removes the empty-output risk above, since nothing is spent on
  reasoning. Use the default for multi-step logic, math, and planning.
  "low" | "medium" | "high" are accepted for OpenAI SDK compatibility but are NOT
  differentiated — all three mean reasoning ON.
- BILLING: credits are charged on total tokens, and reasoning tokens count toward
  completion_tokens even though the reasoning text is never returned. That is why
  reasoning_effort "none" is the single biggest cost lever on this endpoint.
- VISION: a turn's "content" is a string OR an array of parts to send an image:
      "content": [
        { "type": "text", "text": "What card is this?" },
        { "type": "image_url", "image_url": { "url": "https://cdn.example.com/card.jpg" } }
      ]
  image_url.url accepts an https:// URL or a base64 data:image/...;base64,... URL.
  Base64 must fit the ~4.5MB request-body limit — use an https URL for larger images.
Response body:
    {
      "id": "chatcmpl-...", "object": "chat.completion", "created": 0, "model": "qwen",
      "choices": [{ "index": 0, "message": { "role": "assistant", "content": "..." }, "finish_reason": "stop" }],
      "usage": { "prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0 },
      "credits": 1
    }

## POST /api/v1/embeddings  (scope: embeddings)
Request body: { "model": "nomic-embed-text", "input": "text" | ["t1", "t2"] }
- Models: nomic-embed-text (text), clip (text + image, shared space), dinov3 (image).
- input: up to 100 items, each <= 8192 chars.
Response body:
    {
      "model": "nomic-embed-text",
      "data": [{ "index": 0, "embedding": [0.01, -0.02] }],
      "usage": { "total_tokens": 3 },
      "credits": 1
    }

## POST /api/v1/image-embeddings  (scope: embeddings)
DINOv3 image embeddings (1280-dim, L2-normalized). Custom (non-OpenAI) contract.
Request body: one of, up to 60 images per batch (backend is CPU, ~0.3s/image):
    { "urls": ["https://cdn.example.com/card.jpg"] }
    { "images": [{ "b64": "<base64 image>" }] }
- urls must be https://. Do not mix urls and images in one request.
Response body:
    { "model": "dinov3", "dim": 1280,
      "embeddings": [[0.01, -0.02, "..."]],
      "errors": [], "credits": 1 }

## Health checks
- GET /api/v1/status  (no auth)  -> { "status": "ok", "timestamp": "..." }   (liveness)
- GET /api/v1/health  (auth)     -> { "ok": true, "organizationId": "...", "scopes": ["chat"] }

## Errors (always check the HTTP status before parsing the body)
- 400  { "error": { "code": "invalid_request | invalid_json | unknown_model", "message": "..." } }
- 401  { "error": "Missing API key" }  or  { "error": "Invalid or revoked API key" }
- 403  { "error": "API key is not scoped for \"chat\"" }
- 402  { "error": { "code": "quota_exceeded", "message": "..." }, "quota": { "used": 0, "limit": 0, "remaining": 0 } }
- 429  { "error": "...", "code": "RATE_LIMIT_EXCEEDED", "retryAfter": 30 }  plus a Retry-After header
- 502  { "error": { "code": "upstream_error", "message": "..." } }

## Reliability rules
- Every successful response returns "credits" (drawn from the organization AI pool). Budget for it.
- For safe retries, send an Idempotency-Key: <uuid> header on chat and embeddings.
- Requests are rate limited per API key; on 429, honor Retry-After and back off.

## Reference call (Node)
    const res = await fetch("https://api.webanto.com/api/v1/chat", {
      method: "POST",
      headers: {
        Authorization: `Bearer ${process.env.WEBANTO_API_KEY}`,
        "Content-Type": "application/json",
        "Idempotency-Key": crypto.randomUUID(),
      },
      body: JSON.stringify({
        model: "qwen",
        messages: [{ role: "user", content: "Explain embeddings in one sentence." }],
      }),
    });
    if (!res.ok) throw new Error(`Webanto API ${res.status}: ${await res.text()}`);
    const { choices, credits } = await res.json();