StrictJSON is a Python framework designed to enforce structured output from Large Language Models (LLMs), primarily by parsing responses as YAML and validating them against Pydantic schemas. It supports a wide range of data types and offers LLM-based error correction for up to three retries, aiming to streamline LLM interactions for developers and researchers.

How It Works

The core innovation is the parse_yaml function, which treats LLM output as YAML. This approach leverages YAML's conciseness and better handling of code blocks compared to JSON. The framework automatically converts a user-defined output_format into a Pydantic schema, which is then used for validation. This allows for robust data type checking and error correction, ensuring the LLM's output conforms to the desired structure.

Quick Start & Requirements

• Install via pip: pip install strictjson
• Requires Python 3.7+
• Supports various LLMs including OpenAI, Claude, and Llama models.
• Tutorials available: parse_yaml.ipynb, strict_json.ipynb

Highlighted Details

• Supports complex nested data structures (lists, dictionaries, unions, optionals).
• Offers type forcing with LLM-based error correction (e.g., type: int, type: ensure <requirement>).
• Integrates with OpenAI's JSON Mode.
• Provides asynchronous processing capabilities (strict_json_async).

Maintenance & Community

• Project initiated on April 7, 2023.
• Open to collaborators.
• Discussion channel available via Discord: discord.gg/bzp87AHJy5.
• Related project: AgentJo (Agentic Framework).

Licensing & Compatibility

• License: MIT.
• Compatible with commercial use and closed-source applications.

Limitations & Caveats

The README notes that smaller versions of Llama 3.2 and 3.1 models may not work well with YAML parsing. Type checking is not supported when using OpenAI JSON Mode.