Question 1

How close is this to the official JS and Python SDKs - can I port my error handling across?

Accepted Answer

The stated goal is 1:1 feature parity on the same wire contract, environment configuration, retry semantics, request metadata and typed errors, adapted to idiomatic Go, with versions tracking the same release line. It is explicit that it is not created, maintained or endorsed by TypeSafe AI. If parity matters for your logs, note one asymmetry: SystemOneAs returns no metadata on a successful decode, so you need SystemOne when you want the request ID on success.

Question 2

SystemOneAs quietly gives me zero values for missing answers - how do I avoid that trap?

Accepted Answer

Declare pointer fields and nil-check them, or go back to SystemOne. Decoding uses encoding/json, which checks field types but not presence, so a missing answer or field decodes to its zero value with no error. Unknown answer types are also dropped, from both the lifted keys and any retained answers object. If you want to keep the full response alongside lifted fields, declare a nested Answers map[string]json.RawMessage; the full interface map and the typed lookup helpers exist only on SystemOne.

Question 3

Can I send this through an internal gateway instead of api.typesafe.ai?

Accepted Answer

Yes. Set Config.BaseURL, or the TYPESAFE_BASE_URL environment variable, to the gateway; the gateway forwards the Authorization header and other request headers unchanged. Explicit Config values take precedence over environment variables, alongside TYPESAFE_API_KEY and TYPESAFE_DEFAULT_MODEL which defaults to jev-latest. The key is validated at construction: it is trimmed and must be non-empty printable ASCII with no whitespace, otherwise you get ErrInvalidRequest.

Question 4

What exactly gets retried, and can I stop it from retrying?

Accepted Answer

The client retries HTTP 408, 429 and 5xx responses, connection errors and timeouts, twice with exponential backoff, at a default 10-second timeout per attempt. You can change that client-wide through Config.Retry or per call through RequestOptions, and cancellation is context-aware throughout. What the README does not document: whether backoff includes jitter, whether Retry-After is honored, and whether there is a total time budget for a call.

Question 5

What Go version and license does it need, and how are releases run?

Accepted Answer

Go 1.25 or newer, MIT licensed, with CI on Go 1.25 and 1.26. It tracks the same release version line as the official JavaScript and Python SDKs: Version, the newest changelog entry and the Git tag must all match exactly, so release 0.6.0 pairs with tag v0.6.0.

Question 6

How does the client check my API key?

Accepted Answer

At construction. The key is trimmed and must be non-empty printable ASCII with no whitespace; a key containing whitespace, control or non-ASCII characters fails fast with ErrInvalidRequest before any request is attempted. Config values also take precedence over TYPESAFE_API_KEY, TYPESAFE_BASE_URL and TYPESAFE_DEFAULT_MODEL.

Question 7

Can I hit the live API without spending on evaluations?

Accepted Answer

Yes, through the opt-in integration tests, which are read-only and exercise GET /v1/models. They run only with TYPESAFE_RUN_LIVE_TESTS=1 and a key set, separate from the normal go test -race -cover suite. Model discovery is also exposed as client.Models.List(ctx).

TypeSafe SDK Go — valksor

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community