AI Fusion docs

1. Quick start

Three steps to your first request:

1. Sign up for a free tenant.
2. Visit Tokens and create a project token (afp_live_…). Copy it — the plaintext is shown once.
3. Make a request:

curl https://ai.viktorarsov.com/v1/chat/completions \
  -H "Authorization: Bearer afp_live_…" \
  -H "Content-Type: application/json" \
  -d '{"model":"gateway:fast-free","messages":[{"role":"user","content":"hi"}]}'

2. Endpoints

All paths are mounted under the root host. See /docs (FastAPI Swagger) for full request and response schemas.

3. Authentication

Every /v1/* call must include a project token in the standard bearer header:

Authorization: Bearer afp_live_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx

Tokens are issued from Dashboard → Tokens. Only the SHA-256 hash is stored server-side; the plaintext is shown once on creation.

Channel pinning suffix

Append :<channel> to your model string to pin to a routing channel — for example gateway:fast-free:dev isolates a dev workload's quota windows from your prod traffic. Channels are bookkeeping only; they don't change which provider answers.

4. Models & aliases

You can address models three ways:

• Alias — gateway:fast-free. Resolves to a ranked list of (provider, model) pairs.
• Provider-qualified — groq/llama-3.1-8b-instant. Pin to one provider.
• Bare — llama-3.3-70b. Matches across providers by display name.

Aliases

Providers

Models

5. Live progress streaming

POST /v1/chat/completions/progress accepts the same JSON body as /v1/chat/completions but always responds with text/event-stream. Each SSE event has a type field describing what just happened in the spiral:

event: progress
data: {"type":"attempt","provider":"groq","model":"llama-3.1-8b-instant","key":"k_3"}

event: progress
data: {"type":"failure","provider":"groq","status":429,"reason":"rate_limited"}

event: progress
data: {"type":"attempt","provider":"cerebras","model":"llama-3.3-70b","key":"k_1"}

event: progress
data: {"type":"success","provider":"cerebras","latency_ms":612,"tokens":248}

event: done
data: {"choices":[{"message":{"role":"assistant","content":"…"}}]}

Clients should keep the connection open until they see an event: done frame.

6. Cross-format ingress

The same model pool can be reached using non-OpenAI request shapes:

• POST /v1/messages — Anthropic Messages API surface. Same auth header, response shape matches Anthropic's.
• POST /v1beta/models/{model}:generateContent — Google Gemini REST shape. Pass an x-goog-api-key header containing your afp_live_… token, or use ?key=….

Both surfaces translate to the OpenAI-compatible internal pipeline so the same routing, quotas, and webhooks apply.

7. Quotas & rate limits

Three layers cooperate:

• Plan caps — monthly request and token ceilings, set per plan.
• RPM cap — sliding-window per project token, enforced in Redis.
• Per-key daily quotas — enforced atomically inside the routing layer, with cooldown until provider midnight on 429/401.

When all keys are cooling, the gateway parks the request up to max_wait_ms (passed via the x-gateway-max-wait-ms header), then either serves it or returns 429 with a next_slot_eta_s body.

8. Error codes

9. Webhooks

Configure target URLs in Dashboard → Webhooks. We POST a JSON body and an HMAC-SHA256 signature in X-AFP-Signature:

{
  "type": "request.completed",
  "tenant_id": "…",
  "request_id": "…",
  "model": "groq/llama-3.1-8b-instant",
  "tokens_in": 41,
  "tokens_out": 207,
  "cost_usd": 0.0000,
  "savings_usd": 0.00031,
  "latency_ms": 612,
  "ts": "2026-04-21T10:14:33Z"
}

Event types:

• request.completed — every successful chat / embeddings / image call.
• request.failed — every terminal failure after the spiral exhausts.
• quota.threshold — fired at 80% and 100% of the monthly cap.
• invoice.created — emitted at the start of each billing period.

10. Response headers we add

Every successful response includes: