LarchSys.OpenAi.JsonSchema 0.3.1

OpenAi JsonSchema

👁 GitHub Actions Workflow Status
👁 NuGet Version

OpenAi-JsonSchema is a lightweight library for generating valid JSON Schema for OpenAI's Structured Outputs feature, ensuring compatibility with OpenAI's JSON Schema subset. It simplifies the creation of structured outputs for OpenAI models, following the schema generation guidelines provided by OpenAI.

Features

• Supports System.ComponentModel.DescriptionAttribute for generating descriptions in the JSON Schema.
• Handles Nullable<T> types (e.g., int?) and nullable reference types (e.g., string?).
• Supports a wide range of types, including primitives such as bool, int, double, and DateTime.
• Automatically manages $defs and $ref in the schema (e.g., "$ref": "#/$defs/MyType").
• Ensures compatibility with OpenAI's JSON Schema format for structured outputs.
• Polymorphism support using JsonPolymorphicAttribute and JsonDerivedTypeAttribute to generate schemas for polymorphic types.
• Supports JsonIgnoreAttribute and all other System.Text.Json attributes for flexible schema customization.

Static Schema Generation

It supports generating a schema from a class, record, struct:

var resolver = new DefaultSchemaGenerator();
var schema = resolver.Generate<MyModel>(options);

Fluent Schema Generation

It supports generating a dynamic custom schema:

• Enables dynamic descriptions for properties.
• Allows picking properties form a type without including all properties in schema.
• Allows defining a schema for JsonElement, JsonNode or object typed properties.
• And much more..

var generator = new DefaultSchemaGenerator();
var schema1 = generator.Build(new JsonSchemaOptions(SchemaDefaults.OpenAi, Helper.JsonOptionsSnakeCase), _ => _
 .Object("A person", _ => _
 .Property<string>("fullName", "Firstname and Lastname")
 .Property("metaData", "Some Metadata", _ => _
 .Object(_ => _
 .Property("type", _ => _.Const("meta"))
 .Property<string>("author", "Author of the document")
 .Property<DateTime>("published", "published date")
 )
 )
 .Property("extra", "Some Extra", _ => _
 .AnyOf(
 _ => _.Object<Person>(),
 _ => _.Object<Organization>()
 )
 )
 )
);

var schema2 = generator.Build(new JsonSchemaOptions(SchemaDefaults.OpenAi, Helper.JsonOptions), _ => _
 .Object<Document>("A document", _ => _
 .Property(_ => _.Id, "Id of the document")
 .Property(_ => _.Name, "Document Name")
 // Metadata is of type JsonElement:
 .Property(_ => _.Metadata, "Some Metadata", _ => _
 .AnyOf(
 _ => _.Object(_ => _
 .Property("type", _ => _.Const("version-1"))
 .Property<string>("author", "Author of the document")
 .Property<DateTime>("published", "published date")
 ),
 _ => _.Object(_ => _
 .Property("type", _ => _.Const("version-2"))
 .Property<string>("publisher", "Author of the document")
 .Property<DateTime>("created", "created date")
 ),
 )
 )
 )
);

Installation

Install via NuGet:

Install-Package LarchSys.OpenAi.JsonSchema

Quick Start

The following example demonstrates how to generate a JSON Schema using the LarchSys.OpenAi.JsonSchema library.

// use Json Options to control PropertyName and Enum serialization:
var jsonOptions = new JsonSerializerOptions(JsonSerializerDefaults.Web) {
 PropertyNamingPolicy = JsonNamingPolicy.SnakeCaseLower,
 Converters = { new JsonStringEnumConverter(JsonNamingPolicy.SnakeCaseLower) }
};

// use SchemaDefaults.OpenAi to enforce OpenAi rule set:
var options = new JsonSchemaOptions(SchemaDefaults.OpenAi, jsonOptions);

var resolver = new DefaultSchemaGenerator();
var schema = resolver.Generate<Document>(options);

var json = schema.ToJsonNode().ToJsonString(new JsonSerializerOptions() { WriteIndented = true });
output.WriteLine(json);
Assert.NotNull(json);

string response = ChatCompletionWithStructuredOutput(messages: [...], model: "...", schema: schema.ToJsonNode());
Assert.StartsWith("""{"id":""", response);

// deserialize with JsonOptions used for schema generation:
var result = JsonSerializer.Deserialize<Document>(response, jsonOptions);
Assert.NotNull(result);
Assert.NotNull(result.Name);

// Model definition with descriptions to include in schema:

[Description("A document")]
public record Document(
 [property: Description("Id of the document")] int Id,
 [property: Description("Document name")] string Name,
 [property: Description("Text lines of the document")] Line[] Lines,
 [property: Description("Next document in order")] Document? Next,
 [property: Description("Prev document in order")] Document? Prev
);

[Description("A line of text in a document")]
public record Line(
 [property: Description("Line number")] int Number,
 [property: Description("Line text")] string Text
);

Example Output

{
 "type": "object",
 "description": "A document",
 "properties": {
 "id": {
 "type": "integer",
 "description": "Id of the document"
 },
 "name": {
 "type": "string",
 "description": "Document name"
 },
 "lines": {
 "type": "array",
 "description": "Text lines of the document",
 "items": {
 "type": "object",
 "description": "A line of text in a document",
 "properties": {
 "number": {
 "type": "integer",
 "description": "Line number"
 },
 "text": {
 "type": "string",
 "description": "Line text"
 }
 },
 "required": [
 "number",
 "text"
 ],
 "additionalProperties": false
 }
 },
 "next": {
 "description": "Next document in order",
 "anyOf": [
 { "type": "null" },
 { "$ref": "#" }
 ]
 },
 "prev": {
 "description": "Prev document in order",
 "anyOf": [
 { "type": "null" },
 { "$ref": "#" }
 ]
 }
 },
 "required": [
 "id",
 "name",
 "lines",
 "next",
 "prev"
 ],
 "additionalProperties": false
}

How It Works

OpenAi JsonSchema simplifies the generation of JSON Schema for structured outputs using C# classes and attributes. It leverages the System.ComponentModel.DescriptionAttribute for field descriptions and supports nullable reference types, ensuring full compatibility with the JSON Schema language supported by OpenAI models.

For more details on the OpenAI Structured Outputs feature, check out:

• Introducing Structured Outputs in the OpenAI API
• OpenAI Structured Outputs Guide

Contributing

Contributions are welcome! Please fork this repository and submit a pull request with any improvements or feature additions. All contributions should follow the repository's guidelines.

License

This project is licensed under the MIT License. See the file for more information.

Contact

For any questions or issues, feel free to reach out via GitHub or email.