User repository on GitHub
An Elixir client for TypeSafe AI with typed responses and bounded concurrency
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Repos & SDKs
TypeSafe AI — hfiguera
Elixir client for the TypeSafe AI System One API: supervised Mint HTTP/1 and HTTP/2 connections, bounded concurrency, request deadlines, retries and telemetry, with typed Choice, Score and Noul answers.
- Stars
- 0
- Language
- Elixir
- License
- MIT
- Published by
- Community
- Author
- hfiguera
- Last verified
- Oct 2026
This is one of several Jev listings for the same project. The canonical entry is TypeSafe AI — api-evangelist.
About this project
TypeSafe is an Elixir client for the TypeSafe AI System One API, designed for latency-sensitive applications. Evaluate shared state with Choice, Score, and Noul questions, and receive validated, typed answers.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Elixir
Highlights
- Built directly on Mint and Elixir's native JSON, with reusable supervised connections per client.
- Question ids, Choice labels, probability keys and Score legend keys stay strings; no atoms are created from server input.
- Retry defaults to 3 attempts on 429 and 529 with capped exponential backoff and full jitter; Retry-After wins.
- pool_size optionally spreads work across supervised workers; the default remains one connection per client.
Watch out
It is independently maintained, not a TypeSafe-owned SDK, and needs Elixir 1.18 and OTP 27 or newer. The library reads no environment variables or Keychain itself, so you pass TYPESAFE_API_KEY in explicitly and use distinct child spec ids for multiple clients. Custom structs need a native JSON.Encoder; Jason.Encoder does not apply, and Choice answers carry a full probability map while Score's value ranges from zero to the last level's index. Transport failures after submission are ambiguous because the server may already have billed the evaluation, so replay needs retry_transport: true and is off by default. The network slot is released before decoding, so caller concurrency has to be bounded too, clients must run on the calling node, and there is no WebSocket transport. Raising pool_size does not raise your service quota. 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.
Deliberate. Your application retrieves the key and passes it into the child spec, so the library reads no environment variables and no Keychain. That is what lets separate clients hold different keys, each started with its own Supervisor.child_spec id. Connections are established lazily, so nothing dials out until the first evaluation.
A dead caller releases its slot, and cancellation differs by protocol: HTTP/1 cancellation closes the connection, HTTP/2 cancellation resets only that stream. Subsequent work reconnects as needed. The caveat is that socket sends have a one-second upper timeout, so a blocked send or scheduling delay can postpone a timeout result beyond your deadline.
Not necessarily, and the README says to start with one connection and measure before increasing it. More connections do not increase your service quota. Each worker bounds outstanding work by its protocol capacity plus max_queue, before negotiation conservatively treated as one, rejected work tries other workers, and if all are full the call returns :overloaded.
A typed struct. TypeSafe.Answer.Choice carries the selected option plus a full probability map, TypeSafe.Answer.Score carries probabilities and a legend with a value that can fall between levels, and TypeSafe.Answer.Noul is a probability from 0 to 1 with no separate confidence field. Every threshold stays your own decision; the library hands you probabilities, not verdicts.
Elixir 1.18+ and Erlang/OTP 27+, added to mix.exs as {:typesafe_ai, "~> 0.1.1"}. Start a TypeSafe.Client child under your supervisor with a name, the api_key your application fetched, and an optional model; use distinct :id values via Supervisor.child_spec/2 to run several clients with different keys. Connections are established lazily, so nothing dials out until the first evaluation.
base_url defaults to https://api.typesafe.ai and model to jev-latest, with timeout 30_000 ms and connect_timeout 5_000 ms. pool_size is 1 and max_concurrency 10 HTTP/2 streams per connection, max_queue 100 additional outstanding slots, and max_response_bytes 8,388,608. Protocols default to [:http1, :http2] and TLS verifies against OTP's system CA store with no option to disable it.
Choice permits 1 to 255 options and accepts nil descriptions, Score requires 2 to 10 ordered levels, and Noul optionally takes a %{"true" => ..., "false" => ...} map as its second argument. State accepts a string, map, or list of JSON-compatible values, and instructions and criteria descriptions may be structured maps or lists.
More Jev resources like this
1GitHub stars
An unofficial Elixir SDK for TypeSafe's System One endpoint, built on Req and Finch, with per-call retry policies, a shared connection pool and status-mapped errors.
Jev Repos & SDKs#community#jevbest#catalog-import
Community
vinnie357/typesafe_sdk_exElixir
vinnie357/typesafe_sdk_exElixir2GitHub stars
An unofficial Elixir client for TypeSafe's System One API. Noul, Choice and Score questions go out in one request and come back as structs with probabilities, and thresholds and policy stay in your code.
Jev Repos & SDKs#community#jevbest#catalog-import
Community
mattneel/typesafeElixir
mattneel/typesafeElixir0GitHub stars
Unofficial typed Elixir client for the TypeSafe System One API: typed question structs in, typed answer structs out, atom-keyed answers, concurrent fan-out over many states and offline test stubs.
Jev Repos & SDKs#community#jevbest#catalog-import
Community
typesend/typesafe_aiElixir
typesend/typesafe_aiElixir45GitHub 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-jevGo16GitHub stars
Swift library and CLI for asking Jev typed Choice, Noul, and Score questions from Apple platforms, with answers returned alongside calibrated probabilities in one parallel request.
Jev Repos & SDKs#community#jevbest#catalog-import
Community
d-date/swift-jevSwift
d-date/swift-jevSwift9GitHub 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-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.