AI Features

System One (TypeSafe Jev)

Jev, from TypeSafe AI, is a decision model, not a chat model. You send it a state (text or JSON) and a set of typed questions; it answers each one with a calibrated value: a yes/no probability, one option from a list, or a position on a rubric, rather than generated text. VectorAxis serves TypeSafe’s System One API on its own paths and in its own format, so the TypeSafe SDK works by changing its base URL.

python

from typesafe_sdk import Choice, Noul, Score, TypeSafeClient

client = TypeSafeClient(
    api_key="vk-1a2b3c4d5e6f7a8b",         # a TYPESAFE virtual key, not your TypeSafe key
    base_url="https://api.vectoraxis.ai",  # the SDK appends /v1/systemone
)

result = client.system_one(
    {"ticket": "I was charged twice. Please fix this ASAP."},
    {
        "billing": Noul(instructions="Is this ticket about billing?"),
        "tone": Choice(
            instructions="What is the customer's tone?",
            criteria={"calm": None, "frustrated": None, "angry": None},
        ),
        "urgency": Score(
            instructions="How urgent is this ticket?",
            criteria=["can wait", "this week", "today"],
        ),
    },
)

print(result.nouls["billing"].noul)    # probability the answer is yes, 0-1
print(result.choices["tone"].choice)   # always one of your options
print(result.scores["urgency"].score)  # 0 = first level ... 2 = last level

Environment variables work too. Instead of passing base_url, set TYPESAFE_BASE_URL=https://api.vectoraxis.ai and TYPESAFE_API_KEY=vk-…. Both are read by TypeSafe’s SDK; nothing else in your code changes.

When to use it

Classification and triage. Route a ticket, label a document, pick a queue. A choice answer is always one of the options you defined, so your code can switch on it directly.
Checks and scores inside a pipeline. “Does this answer cite the source?”, “how severe is this incident?”, with a confidence you can threshold on to decide between acting, asking for review, or falling back.
Not for writing. Jev doesn’t generate text, code, or summaries. Use chat completions for that.

Setting up a key

Create a virtual key in Bring Your Own Key mode, pick TypeSafe AI as the provider, and paste the API key from your TypeSafe account. Send the key’s slug as Authorization: Bearer vk-… (what the SDK does) or x-virtual-key. For a quick test without a virtual key, send x-provider: typesafe plus x-api-key.

A TypeSafe key only works here. It is rejected with a 422 on chat completions and on every audio, image, file, batch and Responses endpoint, and any other provider’s key is rejected on /v1/systemone. Platform (credit-mode) keys are not supported for System One yet.

Endpoints

Method	Path	What it does
POST	/v1/systemone	Answer every question about the state in one call. Questions run in parallel and independently.
GET	/v1/models	The model names your TypeSafe account can use (what the SDK’s models.list() calls).

Request and response bodies are TypeSafe’s own, unchanged: VectorAxis settings travel in headers, never in the body. The same call over plain HTTP:

bash

curl https://api.vectoraxis.ai/v1/systemone \
  -H "Authorization: Bearer vk-1a2b3c4d5e6f7a8b" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "jev-latest",
    "state": "Help! My payouts have been failing for 3 days.",
    "questions": {
      "is_urgent": { "type": "noul", "instructions": "Does this convey urgency?" }
    }
  }'

json

{
  "model": "jev-1.13.0",
  "answers": { "is_urgent": { "type": "noul", "noul": 0.95 } },
  "usage": { "input_tokens": 296, "output_tokens": 20 }
}

Aliases move. jev-latest points at TypeSafe’s newest stable release; the response’s model is the exact version that answered, and VectorAxis records it in your logs. If you tune thresholds against one version, send its versioned id instead of the alias.

The three question types

type	criteria	Answer
noul	Optional { "true": …, "false": … } describing what yes and no mean.	noul: the probability the answer is yes, 0–1.
choice	A map of option → description (or null). Up to 255 options.	choice, a probabilities map over every option, and confidence.
score	An ordered list of 2–10 level descriptions, lowest first.	score (a position that can land between levels), a legend, probabilities and confidence.

instructions, criteria and the state itself can all be structured JSON as well as plain text. TypeSafe documents the full semantics, including how confidence is derived, in its primitives guide and API reference.

What VectorAxis adds

	How it applies to System One
Budgets & limits	The key’s credit limit, alert threshold and expiry apply, as do its requests-per-minute, per-day and tokens-per-minute limits (every call counts as a request; its input and output tokens count toward TPM), and your plan’s workspace request limit. Spend on System One counts against the same credit limit as everything else on the key, and a key that runs out moves to EXHAUSTED.
Model allowlist	Restrict a key to specific models, e.g. jev-latest or a pinned jev-1.13.0.
Guardrails	Input guardrails (the key’s own and any named in x-guardrail) run over every piece of text in the state and in each question’s instructions. BLOCK stops the call before it reaches TypeSafe; REDACT masks the text first. There is no output phase: answers are numbers and your own option names.
Logs	Every call is a log entry with operation system_one, the version that answered, the number of questions, and the lowest confidence in the call, so the least certain decisions are easy to find. See Request Logs.
Cost	Priced at TypeSafe’s published rate for Jev: $0.042 per million input tokens, output tokens free. The cost of each call is on its log entry and in the x-operation-cost-usd response header.
Retries	429 and 529 from TypeSafe are retried with backoff, honouring Retry-After; tune with x-retry-count / x-retry-status-codes.
Errors	TypeSafe’s own errors come back unchanged (status, body and Retry-After), so the SDK’s error handling works as it does direct. Errors raised by VectorAxis itself, such as an exhausted key or a guardrail block, use the platform’s usual format.

Not supported (yet)

Platform keys. Bring your own TypeSafe key.
Caching and fallback. x-cache-* and x-fallback-* headers are ignored on System One.
The playground. It is built for chat; call System One from code or the SDK.

Getting started: create a TypeSafe AI BYOK virtual key, point the TypeSafe SDK’s base_url at https://api.vectoraxis.ai, and use the key’s slug as the API key. Your decisions then show up in your request logs alongside the rest of your traffic.

VectorAxis is built by GK Labs, a member of NVIDIA Inception.