Question 1

Should I build on an unofficial Go SDK, or wait for TypeSafe to ship one?

Accepted Answer

The project states it exists because TypeSafe has no official Go SDK, is not affiliated with or endorsed by them, and should ideally become unnecessary if an official one lands. It is in beta, implements both /v1/systemone and /v1/models, passes go test against local fixtures with no network, and is checked against the upstream OpenAPI spec committed in the repo. Factor that churn risk into your choice.

Question 2

Will my requests be cached or retried behind my back?

Accepted Answer

Not cached: the in-memory SystemOne cache is opt-in through WithCache, disabled by default, and keyed by the request body. Retries are on by default at 3 for network errors, 429s and 5xxs, honoring Retry-After, and configurable. Non-2xx responses come back as *typesafe.Error with StatusCode, Detail when the API sends one, and the raw Body.

Question 3

Are the docs site and the repository both authoritative, or will they drift?

Accepted Answer

The landing page is short and defers detail to subpages, and the substantive API detail lives in api.md, the runnable examples under examples/, and pkg.go.dev, alongside a Mintlify docs site. Which surface is authoritative is not stated anywhere in the README, and there is no documented versioning policy for the docs, so treat the repository as the source and the site as a convenience.

Question 4

Will adding Braintrust tracing pull extra dependencies into my build?

Accepted Answer

Not in the core path. The integration lives in an unmerged fork branch you install with go get github.com/atharvamhaske/braintrust-sdk-go/trace/contrib/typesafe@feat/typesafe-tracing, and its example has its own go.mod, so braintrust-sdk-go and OpenTelemetry stay out of the core SDK's dependency graph. It routes through WithHTTPClient and matches the span shape of the official Python and JS integrations.

Question 5

How do I install it, and what has to be true before the first call?

Accepted Answer

go get github.com/atharvamhaske/typesafe-sdk-go. You need Go 1.21 or newer and a TypeSafe API key in TYPESAFE_API_KEY, which typesafe.NewClient() reads for you. The base URL defaults to https://api.typesafe.ai and the model to jev-latest, both overridable. The repository is MIT licensed and points issues and pull requests at CONTRIBUTING.md.

Question 6

Which model answers my call, and can I pin one per request?

Accepted Answer

The default is jev-latest, set process-wide with WithModel or TYPESAFE_DEFAULT_MODEL and overridden per call with WithRequestModel, so you can pin a version such as jev-1.13.0 for calibrated thresholds without rebuilding the client. ListModels returns the models available to your account, and every response reports the model that answered plus token usage.

Question 7

What can I run from the repository to see it working end to end?

Accepted Answer

Five runnable programs under examples/ cover the basics and three uses: systemone, listmodels, spamfilter as a Noul spam and moderation classifier, prlabel as a Choice-based PR and diff labeler, and gamemove as a Choice-based game move picker. Run one with TYPESAFE_API_KEY=... go run ./examples/systemone; the README's live run printed the chosen label, the score and the noul probability, then usage: 381 in / 64 out.

typesafe-sdk-go Documentation

About this project

Highlights

Questions about this entry

From the Jev guides

From the Jev AI model community