This package is designed to solve two problems that are especially relevant for JSON output from LLMs:

• Parsing JSON while the response is still streaming
• Correcting hallucinated properties and other schema inconsistencies

Usage

This package exposes one method.

aiJson(jsonString: string, options?: AiJsonOptions): T | null

The basic usage of this method is very similar to JSON.parse(...). You can call aiJson('{ "pi": 3.14 }') to get the object { pi: 3.14 }. In the basic usage, the difference is that it avoids throwing errors and instead returns as much data as it can parse before input ends.

Examples

| Input | JSON.parse | aiJson | | ------------------ | -------------- | -------------- | | "{ "pi": 3.14 }" | { pi: 3.14 } | { pi: 3.14 } | | "{ "pi": 3.14" | Error | { pi: 3.14 } | | "{ "pi":" | Error | { pi: null } | | "{ pi: 3.14 }" | Error | Error |

AiJsonOptions

All options are optional. You can call aiJson(...) without any options if you just want a lax version of JSON.parse.

schema

This is a JSON schema. If you are using ChatGPT functions, you can use the same schema that you've defined for the function parameters as the schema for this parser.

Callbacks will be fired for any places where the parsed value doesn't match the provided schema.

Note: Not all schema rules are validated. Specifically, this package validates the type of each field, and the properties of objects (making sure required fields are present, handling unexpected properties).

onError(error)

This is a catch-all error handling callback. If specified, all types of errors will call onError with information about the error that occurred. The specific error types are also listed below with their specific callback options.

onInvalidType(error)

This is called when a value doesn't match the expected type (or enum) of the schema. The error object has these properties:

| Property | Type | Example | | -------- | ------------- | -------------------- | | type | 'InvalidType' | | | path | Path | ['foo', 0] | | expected | JsonSchema | { type: 'string' } | | value | any | |

onMissingProperty(error)

This is called when an object is missing a required property from the schema. The error object has these properties:

| Property | Type | Example | | -------- | ----------------- | -------------------- | | type | 'MissingProperty' | | | path | Path | ['foo', 0] | | expected | JsonSchema | { type: 'string' } | | object | object | | | property | string | 'name' |

onUnexpectedProperty(error)

This is called when an object has a property that is not in the schema. The error object has these properties:

| Property | Type | Example | | -------- | -------------------- | -------------------- | | type | 'UnexpectedProperty' | | | path | Path | ['foo', 0] | | expected | JsonSchema | { type: 'string' } | | object | object | | | property | string | 'nonsenseName' |

Example

To handle JSON data from a ChatGPT response while it's streaming:

import { aiJson } from "@wits/ai-json";

const json = aiJson(completion, {
  schema: {
    type: "object",
    properties: {
      foo: {
        type: "string",
      },
      bar: {
        type: "number",
      },
      required: ["foo", "bar"],
    },
  },
  onError: (error) => {
    if (error.type === "MissingProperty") {
      switch (error.property) {
        case "foo":
          error.object.foo = "";
          break;
        case "bar":
          error.object.bar = 0;
          break;
        default:
          break;
      }
    } else if (error.type === "UnexpectedProperty") {
      delete error.object[error.property];
    } else if (error.type === "InvalidType") {
      if (error.expected.type === "string") {
        return "";
      } else if (error.expected.type === "number") {
        return 0;
      }
    }
  },
});

// json will have an object with the expected shape
// as soon as the first `{` is present