Data Schemas

It is possible to force the LLM to generate data that conforms to a specific schema. This technique works reasonably well and GenAIScript also provides automatic validation “just in case”.

You will notice that the schema supported by GenAIScript is much simpler than the full-blow JSON schema specification. We recommend using simple schemas to avoid confusing the LLM; then port them to your application specific data format later on.

`defSchema`

Use defSchema to define a JSON/YAML schema for the prompt output.

const schema = defSchema("CITY_SCHEMA", {
    type: "array",
    description: "A list of cities with population and elevation information.",
    items: {
        type: "object",
        description: "A city with population and elevation information.",
        properties: {
            name: { type: "string", description: "The name of the city." },
            population: { type: "number", description: "The population of the city." },
            url: { type: "string", description: "The URL of the city's Wikipedia page." }
        },
        required: ["name", "population", "url"]
    }
})

$`Generate ... using JSON compliant with ${schema}.`

Prompt encoding

Following the “All You Need Is Types” approach from TypeChat, the schema is converted TypeScript types before being injected in the LLM prompt.

// A list of cities with population and elevation information.
type CITY_SCHEMA = Array<{
    // The name of the city.
    name: string,
    // The population of the city.
    population: number,
    // The URL of the city's Wikipedia page.
    url: string,
  }>

You can change this behavior by using the { format: "json" } option.

const schema = defSchema("CITY_SCHEMA", {...}, { format: "json" })

The trace allows you to see the schema source and the rendered prompt.

schema CITY_SCHEMA

source:

{
  "type": "array",
  "description": "A list of cities with population and elevation information.",
  "items": {
    "type": "object",
    "description": "A city with population and elevation information.",
    "properties": {
      "name": {
        "type": "string",
        "description": "The name of the city."
      },
      "population": {
        "type": "number",
        "description": "The population of the city."
      },
      "url": {
        "type": "string",
        "description": "The URL of the city's Wikipedia page."
      }
    },
    "required": [
      "name",
      "population",
      "url"
    ]
  }
}

prompt (rendered as typescript):

// A list of cities with population and elevation information.
type CITY_SCHEMA = Array<{
    // The name of the city.
    name: string,
    // The population of the city.
    population: number,
    // The URL of the city's Wikipedia page.
    url: string,
  }>

Use the schema

Then tell the LLM to use this schema to generate data.

const schema = defSchema(...)
$`Use ${schema} for the JSON schema.`

Validation

When a JSON/YAML payload is generated with the schema identifier, GenAIScript automatically validates the payload against the schema.

Repair

GenAIScript will automatically try to repair the data by issues additional messages back to the LLM with the parsing output.