A Kotlin multiplatform JSON Schema library. Useful for AI and LLMs' tool use (function calling), as it generates JSON Schema for Kotlin @Serializable classes.

This library was created to fulfill the need of agentic AI projects created by xemantic. In particular:

• anthropic-sdk-kotlin - an unofficial Kotlin multiplatform variant of Anthropic SDK.
• claudine - AI Agent build on top of this SDK.

These projects are heavily dependent on tool use (function calling) functionality provided by many Large Language Models. Thanks to xemantic-ai-tool-schema, a Kotlin class, with possible additional constraints, can be automatically instantiated from the JSON tool use input provided by the LLM. This way any manual steps of defining JSON schema for the model are avoided, which reduce a chance for errors in the process, and allows to rapidly develop even complex data structures passed to an AI agent.

In short the xemantic-ai-tool-schema library can generate a JSON Schema from any Kotlin class marked as @Serializable, according to kotlinx.serialization.

In build.gradle.kts add:

plugins {
  kotlin("multiplatform") version "2.1.0" // (or jvm for jvm-only project)
  kotlin("plugin.serialization") version "2.1.0"
}

// ...
dependencies {
  implementation("com.xemantic.ai:xemantic-ai-tool-schema:0.1.4")
}

Then in your code you can define entities like this:

@Serializable
@SerialName("address")
@Title("The full address")
@Description("An address of a person or an organization")
data class Address(
  val street: String,
  val city: String,
  @Description("A postal code not limited to particular country")
  @MinLength(3)
  @MaxLength(10)
  val postalCode: String,
  @Pattern("[a-z]{2}")
  val countryCode: String,
  @Format(StringFormat.EMAIL)
  val email: String? = null
)

And when jsonSchemaOf() function is invoked:

val schema = jsonSchemaOf<Address>()

It will produce a JsonSchema instance, which serializes to:

{
  "type": "object",
  "title": "The full address",
  "description": "An address of a person or an organization",
  "properties": {
    "street": {
      "type": "string"
    },
    "city": {
      "type": "string"
    },
    "postalCode": {
      "type": "string",
      "description": "A postal code not limited to particular country",
      "minLength": 3,
      "maxLength": 10
    },
    "countryCode": {
      "type": "string",
      "pattern": "[a-z]{2}"
    },
    "email": {
      "type": "string",
      "format": "email"
    }
  },
  "required": [
    "street",
    "city",
    "postalCode",
    "countryCode"
  ]
}

And this is the input accepted by Large Language Model APIs like OpenAI API and Anthropic API. When requesting a tool use, these LLMs will send a JSON payload adhering to this schema, therefore immediately deserializable as the original @Serializable Kotlin class.

More details and use cases in the JsonSchemaGeneratorTest.

For JVM-only projects, it is possible to specify java.math.BigDecimal serialization. It will serialize decimal numbers to strings, and add description and pattern properties to generated JSON Schema of a BigDecimal property.

See JavaBigDecimalToSchemaTest for details.

There is an interface called Money defined in the tests of this project. It explains how to define and serialize monetary amounts independently of the underlying decimal number and arithmetics provider.

Clone this repo and then in the project dir:

./gradlew build