User repository on GitHub
Unofficial Go client for TypeSafe's System One API and its model, Jev.
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Repos & SDKs
Jev — anilsenay
An unofficial Go client that keeps Jev answers in your own types, so a choice over your enum yields that enum and the compiler checks the switch you write on it.
- Stars
- 1
- Language
- Go
- License
- MIT
- Published by
- Community
- Author
- anilsenay
- Last verified
- Oct 2026
This is one of several Jev listings for the same project. The canonical entry is Jev — mayank953.
About this project
Ask Jev — the decision model behind TypeSafe's System One API — questions whose answers arrive as your own Go types.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Go
Highlights
- Three constructors, Noul, Choice or OneOf, and Score, return typed answers carrying probabilities for every option.
- Sure(threshold) treats a low-confidence answer as policy rather than an error, so the threshold lives in your code.
- Batched questions share one request, run in parallel, and Run returns model, usage, request id and latency together.
- RawQuestion and Spec send a question kind this package predates and hand the answer back undecoded.
Watch out
This is explicitly unofficial and needs an API key from TypeSafe's early access, so you are tracking a first-party API through a third-party wrapper. Confidence is the API's own peakedness field and the README is emphatic that it is not a calibrated probability of being right; use Probs or the noul's P for that, and note a noul carries no separate confidence at all. Pre-flight validation is deliberately narrow: at least one question, at least two score levels, and no duplicate or empty choice labels, while MaxChoiceOptions and MaxScoreLevels are exported for reference but not enforced so a raised server limit cannot block a valid request. The API itself accepts at most ten score levels. One answer that does not fit its question fails the whole Run with an error wrapping ErrMalformedAnswer, though the other handles stay readable. Listing models needs a provider implementing ModelLister; the default HTTP provider does, and anything else returns ErrNoModelList. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
No, and the README spends a paragraph on the distinction. Confidence is the API's own field describing how peaked the distribution is, not a calibrated probability that the answer is correct. For an actual probability read Probs for the chosen value, or P on a noul. A noul answer has no separate confidence field at all because its probability is the signal, which is also why Sure is the thresholding helper on both noul and choice answers.
New fields and question kinds land upstream before a release here catches up, and the escape hatch is RawQuestion, which sends a Spec exactly as given and returns the answer undecoded so nothing blocks you in the meantime. Extra passes arbitrary fields, though type, instructions and criteria are reserved and cannot be overridden from it. Nothing is validated or narrowed on that path, so the README's advice is to reach for a typed constructor whenever one fits.
Only what it can be sure of: at least one question, at least two score levels, and no duplicate or empty choice labels. Notably it does not enforce its own exported MaxChoiceOptions and MaxScoreLevels constants, deliberately, so that a raised server limit never blocks a valid request. The API's own ceiling still applies at the far end, where a score question accepts at most ten levels.
The Run call does fail, wrapping ErrMalformedAnswer, when an answer does not fit the question it was asked. The other handles in the batch are still readable after that, so a single bad question does not force you to re-send the whole set. This is also the reason batching independent questions about the same state is cheap: they run in parallel on the model and cost one call.
go get github.com/anilsenay/jev, then jev.New() reads TYPESAFE_API_KEY from the environment. It is explicitly an unofficial wrapper, so you need a key from TypeSafe's early access and you are tracking a first-party API through a third-party client that may lag it.
Yes. Instructions, choice option descriptions, score levels and noul criteria are all jev.Content: a plain string, or anything that encodes to a JSON object or array. A nil description is sent as JSON null, and jev.Text renders any of it back for logging.
A weighted position, not a label: Value, Probs, Levels, Legend and Confidence. Because it is probability-weighted it can rest between two levels, and the README argues that in-between reading is the useful part. The worked example returns Value 1.47 with Confidence 0.30 and Nearest returning the has-a-workaround level, which rounding would throw away.
More Jev resources like this
45GitHub stars
Go SDK plus a pipeline-friendly CLI for typed decisions, covering single questions, batched question maps, retries, and line-oriented filtering over stdin.
Jev Repos & SDKs#community#jevbest#catalog-import
Community
mattn/go-jevGo
mattn/go-jevGo9GitHub stars
Go SDK for the TypeSafe API that asks typed Noul, Choice, and Score questions about a state and returns probability distributions with confidence and typed accessors.
Jev Repos & SDKs#community#jevbest#catalog-import
Community
Tangerg/typesafe-sdk-goGo
Tangerg/typesafe-sdk-goGo4GitHub stars
Community Go client for the TypeSafe System One evaluation API with raw structs and a builder that produce the same EvaluationRequest, defaulting to jev-latest.
Jev Repos & SDKs#community#jevbest#catalog-import
Community
kisshan13/typesafe-ai-goGo
kisshan13/typesafe-ai-goGo1GitHub stars
Unofficial, independently maintained Go client for the TypeSafe AI System One API with no third-party dependencies: choice, score and noul questions, typed accessors that error instead of returning zero, and per-call timeout, retry and header overrides.
Jev Repos & SDKs#community#jevbest#catalog-import
Community
cole-gillespie/typesafe-goGo
cole-gillespie/typesafe-goGo2GitHub stars
Community Go client for the TypeSafe AI System One API covering the same surface as the official Python and TypeScript SDKs, using functional options, context cancellation, sentinel errors, `log/slog`, no dependencies outside the standard library, and optional Langfuse tracing.
Jev Repos & SDKs#community#jevbest#catalog-import
Community
fgn/jevgoGo
fgn/jevgoGo1GitHub stars
Unofficial Go SDK for TypeSafe AI's System One API, kept deliberately minimal: one client, a state map, a map of typed questions, and answers read back by question id.
Jev Repos & SDKs#community#jevbest#catalog-import
Community
guillemus/jev-goGo
guillemus/jev-goGoFrom the Jev guides
Original write-ups on what this entry is for.
Closed weights have not stopped the ecosystem from rebuilding the shape in the open. API-compatible servers, retrained heads on small models, and inference-time tricks that skip generation entirely.
Jev Repos & SDKs#open-models#research#local
Read article
From the Jev community
Posts from builders shipping with Jev right now.