LLM Structured Outputs

A plugin for the llm CLI tool that enables structured output generation from language models according to a specified schema. For usability, users can define their desired schema in YAML.

Features

• 🔧 Generate structured outputs according to a simple YAML definition
• 🔄 Support for nested objects and references
• 🤖 Uses OpenAI's native structured output API mode
• 🛡️ Built-in validation using Pydantic

Installation

Assuming you have llm installed already, simply run:

llm install -U llm-structure

Usage

Schema Definition

Create a YAML file defining your schema. Note the use of $ref which allows for lists of objects in this case. This is the same as defining List[People].

# people.yaml

People:
  type: array
  items:
    $ref: "#/Person"
Person:
  name: str
  age: int

Then run the plugin:

llm structure "Chase, 53, and 49 year old Mia went to the market." --schema people.yaml -m gpt-4o

Alternatively, you can use STDIN by specifying - as the prompt:

echo "Chase, 53, and 49 year old Mia went to the market." | llm structure - --schema people.yaml -m gpt-4o

This prints output directly to the console in JSON format:

{
    "items": [
        {
            "name": "Chase",
            "age": 53
        },
        {
            "name": "Mia",
            "age": 49
        }
    ]
}

Supported Models

Currently supports OpenAI models:

• o1-2024-12-17 and later
• gpt-4o-mini-2024-07-18 and later
• gpt-4o-2024-08-06 and later

Integration with Gemini (native) and instructor for the balance of models is coming soon.

Requirements

• Python 3.8+
• llm
• PyYAML
• Pydantic