Skip to main content

TypeSafe

TypeSafe provides decision models, such as Jev, through its System One API. TypeSafeDecisionModel implements the DecisionModel API on top of it.

The same API is also implemented by other servers, for example OpenRouter and inference servers such as SGLang running open models. TypeSafeDecisionModel can be used with any of them by setting the base URL (see below).

note

This integration is experimental and may change in future releases.

Maven Dependency​

<dependency>
    <groupId>dev.langchain4j</groupId>
    <artifactId>langchain4j-typesafe</artifactId>
    <version>1.21.0-beta31</version>
</dependency>

Usage​

DecisionModel decisionModel = TypeSafeDecisionModel.builder()
        .apiKey(System.getenv("TYPESAFE_API_KEY"))
        .modelName("jev-1.13.0")
        .build();

DecisionRequest request = DecisionRequest.builder()
        .input("Help! My payouts have been failing for 3 days.")
        .question("team", ChoiceQuestion.builder()
                .text("Which team should handle this ticket?")
                .option("billing", "Payments, payouts, invoices, refunds")
                .option("support", "Problems using the product")
                .build())
        .question("urgent", YesNoQuestion.of("Does this need attention today?"))
        .build();

DecisionResponse response = decisionModel.decide(request);

ChoiceAnswer team = response.choice("team");
YesNoAnswer urgent = response.yesNo("urgent");

See the Decision Models tutorial for all question types and how to use the answers.

Configuration​

TypeSafeDecisionModel decisionModel = TypeSafeDecisionModel.builder()
        .apiKey(...)              // required for api.typesafe.ai, optional for other servers
        .modelName(...)           // required here or on each request (otherwise decide() fails), e.g. "jev-1.13.0"
        .baseUrl(...)             // defaults to "https://api.typesafe.ai"
        .timeout(...)             // defaults to the HTTP client builder's timeouts, otherwise 15s connect / 60s read
        .maxRetries(...)          // retries after the first attempt; defaults to 2
        .customHeaders(...)       // additional HTTP headers, as a Map or a Supplier<Map>
        .listeners(...)           // DecisionModelListeners, see Observability in the Decision Models tutorial
        .logRequests(...)         // defaults to false
        .logResponses(...)        // defaults to false
        .logger(...)              // an alternate SLF4J logger for requests and responses
        .httpClientBuilder(...)   // see below
        .build();

The available models are listed in the TypeSafe documentation. Aliases such as jev-latest can move to a new version at any time; in production, use a fixed version.

The HTTP client can be customized with httpClientBuilder(...), see Customizable HTTP Client.

Both decide() and decideAsync() are supported.

The model supports the YesNoQuestion, ChoiceQuestion and ScaleQuestion question types. In the TypeSafe documentation, yes/no questions are called "noul" questions, scale questions are called "score" questions, and the input is called the "state". The confidence of choice and scale answers is computed by the server.

Answers are validated against the request: a missing answer, an answer of the wrong type, an option that was not offered, a probability outside of 0 to 1, or a scale mean outside of the levels throws InvalidDecisionResponseException. Levels of a scale answer without a reported probability get a probability of 0.

The input and the questions are sent to the server you configure. When using a hosted service, check its data processing and retention terms, and see Data protection.

Other servers​

Any server that implements the System One API (POST /v1/systemone) can be used by setting the base URL. For example, with OpenRouter:

DecisionModel decisionModel = TypeSafeDecisionModel.builder()
        .baseUrl("https://openrouter.ai/api")
        .apiKey(System.getenv("OPENROUTER_API_KEY"))
        .modelName("typesafe/jev-1.13")
        .build();

Or with a self-hosted server:

DecisionModel decisionModel = TypeSafeDecisionModel.builder()
        .baseUrl("http://localhost:8000")
        .modelName("my-local-model")
        .build();

Self-hosted servers usually need no API key.

TypeSafeDecisionModel is tested with the TypeSafe API and with OpenRouter.

Servers can differ in the limits they apply (for example, the maximum number of options), in how they compute confidence and in the quality of their answers: small models and general-purpose language models served as decision models are usually less accurate and less decisive than dedicated decision models. Check the documentation of the server you use, and evaluate the answers on your own data.

Whatever the base URL, the model reports ModelProvider.TYPESAFE as its provider, for example in listeners and metrics.