llmparse

Parse the JSON your LLM meant to send.

LLMs promise JSON and deliver markdown fences, apologies, single quotes, trailing commas, Python literals — and, when the token budget runs out, JSON that stops mid-string. llmparse is a zero-dependency Python library that extracts the JSON from raw model output and repairs the damage, so json.loads stops being the most fragile line in your pipeline.

from llmparse import parse

parse('Sure! Here you go:\n```json\n{name: "ada", tags: ["x", "y",], ok: True}\n```')
# {'name': 'ada', 'tags': ['x', 'y'], 'ok': True}

parse('{"name": "ada", "tags": ["x", "y"')   # truncated stream
# {'name': 'ada', 'tags': ['x', 'y']}

Install

pip install llmparse-py     # once published to PyPI
pip install git+https://github.com/gptfei/llmparse

Python 3.9+. No third-party dependencies, ever.

What it fixes

Brackets and comment markers inside string values are never touched.

API

• parse(text, default=...) — return the recovered Python object. Raises LLMParseError when nothing is recoverable, unless default is given. loads is an alias.
• extract(text) — return the first recoverable JSON substring, or None.
• repair(text) — return repaired JSON text (for when you need the string, not the object).

Valid JSON is returned untouched: every repair rule is a no-op on input that already parses.

Contributing

The repair pipeline is a list of small, independent rules — adding one is the easiest meaningful contribution in open source right now:

1. Find a way an LLM broke JSON that llmparse does not fix yet.
2. Write the rule in src/llmparse/repair.py, register it in RULES.
3. Add a test with the broken input.

See CONTRIBUTING.md and issues labelled good first issue.

中文简介

大模型说好返回 JSON，实际经常带 markdown 围栏、解释文字、单引号、 尾随逗号、Python 的 True/None，甚至输出到一半被截断。llmparse 是一个零依赖的 Python 库：从原始输出里提取 JSON、逐项修复这些损坏、 再解析成 Python 对象。合法 JSON 不会被改动。

License

MIT — see LICENSE.