api.saina.run

Endpoints

Both POST routes take a bearer token: Authorization: Bearer $SAINA_API_KEY. Each call is charged in credits for its input tokens. Errors have a stable code and say whether a retry is safe; send an Idempotency-Key to retry without being charged twice. Details: API docs.

Examples

Native contract

curl https://api.saina.run/v1/ask \
  -H "Authorization: Bearer $SAINA_API_KEY" -H 'Content-Type: application/json' \
  -d '{
    "model": "saina-helm-2-0.8b",
    "mode": "decision", "threshold": 0.8, "min_margin": 0.05,
    "state": "I was charged twice for last month and nobody answers the chat.",
    "questions": {
      "team":     {"type": "single_choice", "question": "Which team should handle this?",
                   "options": {"billing": "Payments and refunds", "technical": "Bugs and outages", "other": null}},
      "tags":     {"type": "multi_choice", "question": "Which tags apply?",
                   "options": {"refund": null, "duplicate_charge": null, "angry": null, "outage": null}},
      "priority": {"type": "rating", "question": "How urgent is this?", "levels": ["low", "normal", "high"]}
    }
  }'

Every answer carries the probability of each option and a reason: accepted, below_threshold, below_margin or tie. An accepted answer clears your configured thresholds; it is not a guarantee of correctness. Apply your own checks and any required human review before acting, and route uncertain answers to a fallback.

System One format

curl https://api.saina.run/v1/systemone \
  -H "Authorization: Bearer $SAINA_API_KEY" -H 'Content-Type: application/json' \
  -d '{
    "model": "saina-helm-2-0.8b",
    "state": "Help! My payouts have been failing for 3 days.",
    "questions": {
      "is_urgent": {"type": "noul",   "instructions": "Does this message convey urgency?"},
      "team":      {"type": "choice", "instructions": "Which team?", "criteria": {"billing": "Money", "technical": "Bugs"}},
      "severity":  {"type": "score",  "instructions": "How severe?", "criteria": ["low", "medium", "high"]}
    }
  }'

Python

from saina import Saina
client = Saina('https://api.saina.run', api_key='YOUR_KEY')
result = client.ask(model='saina-helm-2-0.8b', state='I was charged twice.', questions={
    'issue': {'type': 'single_choice', 'question': 'Identify the issue.',
              'options': {'duplicate': 'duplicate charge', 'delivery': 'card delivery'}}})

The console playground (after you create a free account) builds these requests for you and copies them as curl, Python, JavaScript or n8n. Credits, keys, limits, retries and error codes are in the API docs.

Data policy: zero retention

This policy describes inference-content handling by the hosted API today. Website, waitlist, support and browser storage are described in section 4 of the terms. Material changes follow the notice and acceptance process; changing this page's date alone does not change an existing customer's agreement. Paid-plan retention described in the terms is not enabled by this policy.

• Request and response bodies are never stored. The server keeps nothing between requests: no database, no disk writes, no cache of your text, no analytics on content. Inference runs in memory and the result is returned.
• No per-request logs on our side. The server runs with access logging off. It logs startup and errors only, and those lines never include request content. Error responses name the field that failed, not its value.
• Not used for training. Nothing you send is used to train, evaluate or tune any model.
• Transit. api.saina.run is reached through Cloudflare, which terminates TLS and forwards the request to our server over a private tunnel. Cloudflare sees requests in transit and keeps its own edge metadata (timestamps, status codes, request counts) under its policies; it does not store bodies for us, and we have not enabled any Cloudflare feature that logs them. If you need no intermediary at all, self-host: it's the same server.
• Keys. A key identifies a caller and can be revoked individually. We don't tie keys to request content, because there is no request content kept to tie them to.
• Location. Requests are processed in Canada (British Columbia). Any standby capacity runs under the same policy, and its region is shown on this page while in use. Weights are the public Saina Helm 2 0.8B release, pinned to the revision reported by /healthz, and are only changed deliberately.
• Through OpenRouter. Requests made through OpenRouter pass through OpenRouter under its data policy before reaching this server, where the rules above apply.

Questions about any of this: [email protected]. Use of the hosted API is under the terms of service.

Limits and support

• Up to 256 questions per request. Questions are scored in groups of up to 8 that share one read of the context; each group (context, questions and options) must fit in 8,192 tokens. Longer inputs are rejected, never truncated, and are not charged.
• You pay for the input tokens you send: the context once per request, plus each question and its options. Prompt formatting and output are not charged. Details are in the terms.
• For guaranteed capacity or an on-premise deployment, talk to us.
• You can self-host the published server with the same request contract: run-saina/deploy (Docker) or pip install 'saina[local,server]'. Hardware, model versions, capacity and operational behavior may differ.