Typesafe Ai

Ask TypeSafe's System One models (Jev) typed questions from swamp and branch workflows on the calibrated answers. Jev does not generate text: it takes application state plus typed questions and returns a probability for a yes/no question (Noul), a selected option with a full distribution (Choice), or a weighted position along an ordered rubric (Score).

Every method calls POST /v1/systemone directly with fetch (no SDK), retries transient failures with exponential backoff that honours Retry-After, and writes one resource per evaluation so later workflow steps can read the answer with CEL.

Methods

Each evaluation lands as <spec>-<name> (default name latest) so several evaluations can live side by side, and every resource records model, state, usage, and evaluatedAt.

Credentials

Supply the API key with the apiKey global argument, ideally as a vault reference: ${{ vault.get("typesafe", "TYPESAFE_API_KEY") }}. When omitted the model falls back to the TYPESAFE_API_KEY environment variable. The api-key-configured pre-flight check (label policy) fails early when no key is available.

Usage

swamp extension pull @swamp/typesafe-ai

swamp model create @swamp/typesafe-ai jev \
  --global-arg 'apiKey=${{ vault.get("typesafe", "TYPESAFE_API_KEY") }}'

swamp model method run jev noul \
  --input state="Help! My payouts have been failing for 3 days." \
  --input instructions="Does this convey urgency?" \
  --input name=urgency

# In a workflow guard:
#   ${{ data.latest("jev", "noul-urgency").attributes.noul > 0.8 }}

See the bundled README for the full argument reference and worked examples.