Introduction
Sieve decision models behind a TypeSafe-compatible API. Send content and typed questions, get calibrated probabilities back.
Sieve is a family of decision models. You send a piece of content (the state) and typed questions about it; Sieve reads the state once and returns a probability for every option. It never writes text.
The hosted API at https://api.sthanika.ai speaks TypeSafe's System One contract, so the TypeSafe Python and JavaScript SDKs work once you change the base URL and the key. An OpenAI-style route, POST /v1/decisions, is also available.
Quickstart in 60 seconds
Create an API key
Sign in to the console, open API keys and create a key. Copy it straight away: the full key is shown once.
Point the TypeSafe SDK at Sieve
export TYPESAFE_API_KEY="sieve_app_live_..."
export TYPESAFE_BASE_URL="https://api.sthanika.ai"Both TypeSafe SDKs read these variables, so the code below has no key or URL in it.
Ask a question
pip install "typesafe-sdk[zstd]"from typesafe_sdk import Noul, TypeSafeClient
with TypeSafeClient(model="sieve-latest") as client:
response = client.system_one(
state="I was charged twice. Please fix this ASAP.",
questions={"billing": Noul(instructions="Is this message about billing?")},
)
print(response.nouls["billing"].noul) # probability of yes, for example 0.98Always name a model
The TypeSafe SDKs fall back to jev-latest when no model is set, and the hosted API does not accept that name. Pass sieve-latest (or another name from GET /v1/models) as shown, or set TYPESAFE_DEFAULT_MODEL.
The full quickstart covers sign-up, reading the response and compression.
Models
| Name | What it is |
|---|---|
sieve-latest | The latest generally available model (currently sieve-9b-v2). A good default. |
sieve-auto | A cascade: Sieve-4B answers first and Sieve-31B is asked again for low-confidence questions. See Auto cascade. |
sieve-preview | A preview model (currently sieve-27b-v2). It may change without notice. |
sieve-0.8b, sieve-2b, sieve-4b, sieve-9b, sieve-27b, sieve-31b | One name per model size. Each points to the current version of that size. |
Every response reports the versioned model that answered, such as sieve-9b-v2, never the alias. GET /v1/models lists the names your key may use. See the model cards for each model.
Explore the docs
Get started
Sign up, create a key and make your first decision.
Question types
Yes/no, choice and score questions, and what each returns.
Route uncertain answers
Use confidence to send doubtful answers to a person.
API reference
Endpoints, headers, limits and every error body.
SDKs
Configure the TypeSafe Python and JavaScript SDKs for Sieve.
Migrate from Jev
Move from TypeSafe (Jev) by changing the base URL and the key.
Enterprise
Vault, Foundry, Assurance and regions.
Billing
Prices, prepaid credits, invoices and tax.
Policies
What we keep, security, acceptable use and legal terms.
About these docs
- This site sets no cookies and loads no analytics. Search runs in your browser. The API reference's Try it sends requests only to the sandbox or a gateway on your machine, and the key you paste stays in the tab's memory.
- Prices on these pages are modelled and will change after benchmarking. See Pricing.
- Questions or corrections: write to support@sthanika.ai.
Last updated