Questions

A library for typed semantic decisions: define questions about your data with Zod 4 schemas or native typed question batches, then inspect evidence and decide what happens next through ordinary async/await or native Web Streams, with Jev via TypeSafe AI or Vercel evaluation.

Install with bun add @nitoba/questions zod (or npm install), create a client with Questions.create({ model: TypeSafe.create({ apiKey, timeout: "15 seconds" }) }), then await questions.about(ticket).is("Is production work blocked?") for an inferred boolean.

Declare questions once, then apply ordered rules to their evidence. The README's release example uses thresholds accept 0.8 and reject 0.2, .when rules producing labels such as block or migration-required, .onUncertain("human-review") and .otherwise("ship") as the fallback.

No. The optional AI SDK generative adapter can use Gemini, Claude, GPT and Gateway language models without rewriting your questions. The core model is Jev through TypeSafe AI or Vercel evaluation, and live model calls still require your own provider credentials.

No - the README is explicit: no Effect dependency, service container or alternative execution runtime. The package was independently inspired by effect-questions but uses ordinary async/await and native Web Streams, and the root import requires the Zod 4 peer even for native batches.

No. The README states decisions are finite classifications, not arbitrary JSON generation, factual guarantees or authorization to execute business actions. Thresholds refer to probability mass rather than provider-specific confidence, and the example values are not universal safety guarantees.

0.1.0 is the stable release on the npm latest tag, including ordered policies and descriptive routing, and the migration guide should be read before upgrading. Each operation invocation starts new work; re-awaiting the same Promise does not rerun it, and construction performs no inference.