repairjson 0.1.2

Rust-backed repair of malformed JSON for LLM-style outputs

repairjson

Blazing-fast JSON repair for messy LLM output.

repairjson is a Rust-powered drop-in repair layer for malformed JSON in Python. It is built for the reality of modern LLM pipelines: broken commas, single quotes, unquoted keys, truncated payloads, Markdown fences, and half-finished responses that still need to get through production systems.

If json.loads() is too strict and Python-side repair is too slow, this is the fast path.

Why repairjson

• Rust core with a single-pass repair engine
• designed for malformed LLM-style JSON, not just clean parser input
• published PyPI wheels
• benchmarked at roughly ~200x average speedup versus Python json_repair on the current 20 MB malformed corpus suite

What It Repairs

• single-pass byte-oriented repair
• support for single-quoted strings
• support for Python literals like True, False, and None
• repair of unquoted object keys
• repair of missing commas and trailing commas
• auto-closing of truncated objects and arrays
• stripping of leading and trailing Markdown code fences
• skipping chatty preambles to recover the first JSON object or array payload

Install

uv add repairjson

For a plain virtual environment:

uv venv
VIRTUAL_ENV=.venv uv pip install --python .venv/bin/python repairjson

Fast Example

uv run --with repairjson python -c "import repairjson; print(repairjson.repair(\"{user: 'alice', active: True, tags: ['x', 'y',],}\"))"

Output:

{"user":"alice","active":true,"tags":["x","y"]}

Usage

import repairjson

fixed = repairjson.repair("{user: 'alice', active: True, tags: ['x', 'y',],}")
print(fixed)
# {"user":"alice","active":true,"tags":["x","y"]}

obj = repairjson.loads("{user: 'alice', active: True, tags: ['x', 'y',],}")
print(obj)
# {'user': 'alice', 'active': True, 'tags': ['x', 'y']}

Performance

Current benchmark suite: six synthetic 20 MB malformed-JSON datasets modeled after LLM-style output patterns.

• dense_object: 206.2x
• fenced_payload: 202.81x
• chatty_nested: 242.24x
• long_text: 198.91x
• array_heavy: 209.86x
• truncated_nested: 229.48x

Average across the current suite: about 214.9x.

The conservative claim is still 100x+, because real-world speedups depend on payload shape, string density, and how broken the JSON is.

The benchmark harness lives in benchmark.py.

Development

Create the local environment and install the package in editable mode:

uv venv
VIRTUAL_ENV=.venv uv pip install --python .venv/bin/python maturin pytest json_repair
uv run maturin develop

Run tests:

cargo test
uv run pytest -q

Run benchmarks:

uv run python benchmark.py --dataset all

Contributing

See CONTRIBUTING.md for the development workflow and pull request expectations.

License

MIT. See LICENSE.