typesafe-jev

Ruby client for Jev, TypeSafe's System One model. Ask natural-language questions about your application state and get typed, probabilistic answers you can program against.

Zero runtime dependencies: pure Ruby stdlib (Net::HTTP, JSON), so it works in plain Ruby and Rails alike.

Installation

gem install typesafe-jev

or in your Gemfile:

gem "typesafe-jev"

Quickstart

require "typesafe"

response = Typesafe::Jev.evaluate(
  state: { ticket: { text: "My payouts failed twice this week." } },
  questions: {
    refund_requested: Typesafe::Noul.new("Does the customer request a refund?"),
    category: Typesafe::Choice.new("Which category does this ticket belong to?", criteria: {
      billing:   "Payment, invoices, refunds",
      technical: "Bugs and outages",
      sales:     "Pre-sales questions"
    }),
    severity: Typesafe::Score.new("How severe is this ticket?", criteria: [
      "Low: routine request",
      "Medium: degraded experience",
      "High: blocked user",
      "Critical: money lost or security issue"
    ])
  }
)

response[:refund_requested].noul  # => 0.95 (yes/no probability, 0..1)
response[:category].choice        # => "billing"
response[:category].probabilities # => { "billing" => 0.88, "technical" => 0.12, "sales" => 0.0 }
response[:severity].score         # => "High: blocked user"
response.usage.input_tokens       # => 296

Typesafe::Jev.evaluate is the one-shot facade with the model pinned to jev-latest. For another model, use Typesafe::Client directly.

Questions

Questions are frozen, immutable value objects. The question id is not in the object; it is the Hash key in the questions: map, and answers come back under the same key.

Class	Answers with	Criteria
Typesafe::Noul	a yes/no probability (noul, 0..1)	optional { true: ..., false: ... }
Typesafe::Choice	one option + full probability distribution + confidence	non-empty { key: "description" }
Typesafe::Score	a level + legend + distribution + confidence	non-empty ["Level 1", "Level 2", ...]

Typesafe::Noul.new("Does the customer request a refund?",
  criteria: { true: "The customer explicitly asks for their money back",
              false: "The customer does not mention refunds" })

Typesafe::Choice.new("Which category does this ticket belong to?",
  criteria: { billing: "Payment, invoices, refunds", technical: "Bugs and outages" })

Typesafe::Score.new("How severe is this ticket?",
  criteria: ["Low", "Medium", "High", "Critical"])

Invalid input raises ArgumentError at construction time, so bad questions fail fast, before any HTTP call.

Responses

Client#evaluate returns a frozen Typesafe::Response:

• response[id] (or response.answers): a typed NoulAnswer, ChoiceAnswer or ScoreAnswer; String and Symbol ids both work
• response.model: the model that performed the evaluation
• response.usage: Typesafe::Usage with input_tokens / output_tokens
• response.to_h / response.to_json: the raw API shape

Each answer type exposes its own accessors (#noul, #choice, #probabilities, #confidence, #score, #legend, …) and serializes via #to_h/#to_json.

Errors

Any non-2xx HTTP response raises a typed error. All inherit from Typesafe::APIError (root: Typesafe::Error), which carries #status, #body, #headers, #request_id and #retryable?.

Error	Status	Extras
Typesafe::BadRequestError	400
Typesafe::AuthenticationError	401
Typesafe::PermissionDeniedError	403
Typesafe::NotFoundError	404
Typesafe::UnprocessableEntityError	422	#errors (validation details)
Typesafe::RateLimitError	429	#retry_after
Typesafe::OverloadedError	529
Typesafe::ServerError	5xx

begin
  Typesafe::Jev.evaluate(state: state, questions: questions)
rescue Typesafe::RateLimitError => e
  sleep(e.retry_after || 1.0)
  retry
rescue Typesafe::OverloadedError
  sleep(2**attempt)
  retry
rescue Typesafe::APIError => e
  raise "typesafe request failed (#{e.status}, request #{e.request_id}): #{e.body}"
end

Configuration

The API key is read from the TYPESAFE_API_KEY environment variable unless passed explicitly:

jev = Typesafe::Jev.new                      # key from ENV
jev = Typesafe::Jev.new(api_key: "sk-...")   # or explicit

client = Typesafe::Client.new(api_key: "sk-...", model: "other-model")
client.evaluate(state:, questions:, model: "one-off-model") # per-call override

Documentation

TypeSafe docs live at docs.typesafe.ai. The full API reference for this gem is on rubydoc.info.

Development

bundle install
bundle exec rake spec   # or: bundle exec rspec

The test suite runs against Ruby 3.1 and 3.4 in CI. See CONTRIBUTING.md for conventions.

License

MIT