Question 1

Does this block my GenServer while a decision is in flight?

Accepted Answer

No. You reply with questions from a callback and the answer arrives as a message your handle_answer/3 clause pattern-matches on, so a hundred calls can be in flight at once and each answer finds its clause when it lands. The request runs under a Task.Supervisor, so a crash becomes {:error, reason} in your callback instead of taking the process down.

Question 2

How do I test code that talks to a model?

Accepted Answer

Nothing below handle_answer/3 touches the network, so tests call it with a literal map. For the transport, point the client at a Req.Test plug and answer with Jev.Test, writing the response the same way you read a reply, with retry_delay: 0. Confidence, probabilities, usage, and model are reserved question names, so your own questions cannot accidentally shadow them.

Question 3

Can I point this at a local model instead of paying per call?

Accepted Answer

Yes, without a key at all. The /v1/systemone wire format is treated as the common shape for decision models, and named endpoints such as laya, kev, decider, and jeff each carry their own base_url, api_key, model, and price, selectable per call or as the default. A named endpoint never inherits the TypeSafe key or its price, which is how a development environment runs keyless.

Question 4

Do my thresholds survive switching to a different model?

Accepted Answer

Treat them as a starting point. Replies have the same shape from any endpoint, and when a server omits confidence it is computed from the probabilities the way TypeSafe defines it, but the README notes models are calibrated differently, so a threshold tuned on one is only a starting point on another. The [:jev, :answer] telemetry, a confidence distribution tagged by question name, works as a calibration monitor.

Question 5

How do I install it, and what does it need?

Accepted Answer

It is on Hex as jev, added with {:jev, "~> 0.1"} and configured with an api_key or left to read TYPESAFE_API_KEY from the environment. Elixir 1.18 or later is required for the built-in JSON module, and Erlang/OTP 27 or later. There is no server process to run; the GenServer is something you define in your own supervision tree.

Question 6

What happens when the call fails?

Accepted Answer

The request runs under a Task.Supervisor, so a crashed request becomes {:error, reason} in handle_answer/3 instead of taking the server down, and a clause matching {:error, reason} treats it like any other reply. At the transport layer 429 and 529 plus gateway 502, 503 and 504 are retried with backoff honouring Retry-After; non-2xx becomes {:error, %Jev.Error{status, body, request_id}}.

Question 7

How do I watch calibration in production?

Accepted Answer

Four telemetry events carry it: [:jev, :request, :start] and :stop with system_time, then duration, input_tokens, output_tokens and cost; [:jev, :request, :exception] with duration; and [:jev, :answer] with confidence and probability tagged by name, type and answer. The state itself never appears in metadata, only its hash. Cost is input tokens at 0.042 per million with no output charge.

Jev for Elixir — dannote

About this project

Highlights

Quickstart

Questions about this entry

Reactions & coverage

Other listings of this project

From the Jev guides

From the Jev AI model community