Question 1

Is an alpha Elixir SDK with a moving model alias safe to build a service on?

Accepted Answer

Two things to weigh. The dependency is 0.1.0-alpha.3, so expect churn. On the model side, jev-latest is an alias TypeSafe can move to a newer version; the client reports whatever versioned id the API resolved, and models/1 lists the ids your account can pin to instead. What does not move: question ids and Choice options come back exactly as you sent them, atoms stay atoms, and nothing in the response is ever converted into an atom.

Question 2

Do I get exceptions on a bad key or a rate limit, or do I have to pattern match?

Accepted Answer

Pattern matching. Every call returns {:ok, result} or {:error, %TypeSafeAPI.Error{}} with a type you can match on: :auth, :validation, :rate_limited, :overloaded, :server_error, :timeout, :connection or :unexpected, and retryable?/1 separates the transient ones. Bang variants such as evaluate!/4 raise the same struct. request_id is carried on both the result and the error, and it is the value to quote when you contact TypeSafe support.

Question 3

Will a failed item in a batch of 5,000 tickets abort the whole run?

Accepted Answer

Not by default. evaluate_many/4 supervises and unlinks its tasks, so a state that raises is one failed outcome rather than a dead caller, and on_error defaults to :collect, returning {:error, _} in place for that state; pass on_error: :raise to fail fast. Two timeouts apply separately: timeout bounds a single HTTP attempt as in evaluate/4, and task_timeout bounds one state's whole task, so it must cover every retry of it.

Question 4

Can my test suite run without network calls or an API key?

Accepted Answer

Yes, via TypeSafeAPI.Test, which stubs the API by question id with no fixtures and decodes into the same structs as a real response. It needs plug in your test dependencies because it builds responses with Req.Test and Plug.Conn. Two catches: a stubbed confidence has to beat the uniform baseline of 1 / n, or the stub decodes to a different option, and any question you do not stub returns HTTP 422, which evaluate/4 turns into a :validation error and evaluate_many/4 turns into one error per state.

Question 5

Do I need to run the client under a supervisor, or can I build it in a function?

Accepted Answer

Build it in a function. new/1 returns a plain struct, not a process, so there is nothing to start under a supervisor. Do not call it in a module attribute, because module attributes are evaluated at compile time before config/runtime.exs or release environment variables load; read the API key at runtime instead.

Question 6

How do I stop the model from moving under me when jev-latest changes?

Accepted Answer

Pin a versioned id. jev-latest is a pinned alias, not a model id, and the API resolves it to something like jev-1.13.0, which is what result.model reports. models/1 lists the model ids your account can pin to instead of the alias.

Question 7

What does each request report about retries and token use?

Accepted Answer

Every request runs inside a [:typesafe_api, :request, :start | :stop | :exception] telemetry span. The :stop event adds status, retry_count, input_tokens, output_tokens and error; attach_logger/1 logs one line per request. An API failure arrives as a :stop event with error set, not an :exception, which only fires when the call itself raises.

TypeSafe AI — typesend

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community