llm-json 0.1.5

A JSON library wrapper that gracefully handles markdown-wrapped JSON from LLMs.

llm-json - stop begging LLMs for valid JSON

llm-json is a Python package that acts as a drop-in replacement for the standard \json\ library. It automatically handles JSON strings that are often returned by Large Language Models (LLMs) wrapped in Markdown code fences or backticks—so you don’t have to beg the AI to always produce “valid JSON.”

Why This Project?

After repeatedly struggling with LLMs returning JSON (but also including backticks, code fences, or extraneous text), I got fed up. Structured output is nice, but it’s not always possible (or desirable) to force the LLM to output perfect JSON. Rather than overcomplicating prompt engineering, llm-json does the messy cleanup for you.

Installation

Grab the latest release from PyPI:

uv pip install llm-json

Can't you just use Structured Outputs?

Yes, but not all providers support them in the same way

Why not use Instructor / Outlines

I love both of those tools, this is just a simple wrapper to make my life easier, I don't need a full blown library for structured outputs, just a simple json validator that works the way LLMs do!

Usage

The core idea is that llm-json works exactly like the standard Python json library, but with automatic fallback parsing for those pesky backticks:

from llm_json import json

# A string returned by an LLM, might be in triple-backtick code blocks or inline backticks
s = """
```json
{
  "hello": "world"
}
```"""

data = json.loads(s)
print(data)  # => {'hello': 'world'}

All other functions (json.dump, json.dumps, json.load) also work the same way. Just replace your existing import json with from llm_json import json, and you’re good to go.

Handling Extra Text

llm-json will ignore extraneous text before and after the code fences and attempt to parse the first valid snippet it finds. It also handles inline backticks (like ```{"hello":"world"}```).

Contributing

Pull requests and issues are very welcome! Check out the GitHub repo to file issues or propose changes.

Testing

We use pytest. To run tests locally:

git clone https://github.com/altryne/llm-json.git
cd llm-json
pip install -e .
pip install pytest
pytest tests

You should see something like:

============================= test session starts =============================
platform darwin -- Python 3.x, pytest-8.x, ...
collected 9 items

tests/test_json.py .........                                                 [100%]

============================== 9 passed in 0.02s ==============================

License

This project is licensed under the MIT License.

Author

Made with ❤️ by @altryne.

• GitHub: github.com/altryne
• X (formerly Twitter): x.com/altryne