This simple package can be used to fix an invalid json string. To know all cases in which this package will work, check out the unit test.

If you find this library useful, you can help me by donating toward my monthly beer budget here: https://github.com/sponsors/mangiucugna

If you are unsure if this library will fix your specific problem, or simply want your json validated online, you can visit the demo site on GitHub pages: https://mangiucugna.github.io/json_repair/

Or hear an audio deepdive generate by Google's NotebookLM for an introduction to the module

Some LLMs are a bit iffy when it comes to returning well formed JSON data, sometimes they skip a parentheses and sometimes they add some words in it, because that's what an LLM does. Luckily, the mistakes LLMs make are simple enough to be fixed without destroying the content.

I searched for a lightweight python package that was able to reliably fix this problem but couldn't find any.

So I wrote one

As part of my job we use OpenAI APIs and we noticed that even with structured output sometimes the result isn't a fully valid json. So we still use this library to cover those outliers.

• Missing quotes, misplaced commas, unescaped characters, and incomplete key-value pairs.
• Missing quotation marks, improperly formatted values (true, false, null), and repairs corrupted key-value structures.

• Incomplete or broken arrays/objects by adding necessary elements (e.g., commas, brackets) or default values (null, "").
• The library can process JSON that includes extra non-JSON characters like comments or improperly placed characters, cleaning them up while maintaining valid structure.

• Automatically completes missing values in JSON fields with reasonable defaults (like empty strings or null), ensuring validity.

Install the library with pip

pip install json-repair

then you can use use it in your code like this

from json_repair import repair_json

good_json_string = repair_json(bad_json_string)
# If the string was super broken this will return an empty string

You can use this library to completely replace json.loads():

import json_repair

decoded_object = json_repair.loads(json_string)

or just

import json_repair

decoded_object = json_repair.repair_json(json_string, return_objects=True)

Some users of this library adopt the following pattern:

obj = {}
try:
    obj = json.loads(string)
except json.JSONDecodeError as e:
    obj = json_repair.loads(string)
    ...

This is wasteful because json_repair will already verify for you if the JSON is valid, if you still want to do that then add skip_json_loads=True to the call as explained the section below.

JSON repair provides also a drop-in replacement for json.load():

import json_repair

try:
    file_descriptor = open(fname, 'rb')
except OSError:
    ...

with file_descriptor:
    decoded_object = json_repair.load(file_descriptor)

and another method to read from a file:

import json_repair

try:
    decoded_object = json_repair.from_file(json_file)
except OSError:
    ...
except IOError:
    ...

Keep in mind that the library will not catch any IO-related exception and those will need to be managed by you

When working with non-Latin characters (such as Chinese, Japanese, or Korean), you need to pass ensure_ascii=False to repair_json() in order to preserve the non-Latin characters in the output.

Here's an example using Chinese characters:

repair_json("{'test_chinese_ascii':'统一码'}")

will return

{"test_chinese_ascii": "\u7edf\u4e00\u7801"}

Instead passing ensure_ascii=False:

repair_json("{'test_chinese_ascii':'统一码'}", ensure_ascii=False)

will return

{"test_chinese_ascii": "统一码"}

If you find this library too slow because is using json.loads() you can skip that by passing skip_json_loads=True to repair_json. Like:

from json_repair import repair_json

good_json_string = repair_json(bad_json_string, skip_json_loads=True)

I made a choice of not using any fast json library to avoid having any external dependency, so that anybody can use it regardless of their stack.

Some rules of thumb to use:

• Setting return_objects=True will always be faster because the parser returns an object already and it doesn't have serialize that object to JSON
• skip_json_loads is faster only if you 100% know that the string is not a valid JSON
• If you are having issues with escaping pass the string as raw string like: r"string with escaping\""

Install the library for command-line with:

pipx install json-repair

to know all options available:

$ json_repair -h
usage: json_repair [-h] [-i] [-o TARGET] [--ensure_ascii] [--indent INDENT] filename

Repair and parse JSON files.

positional arguments:
  filename              The JSON file to repair

options:
  -h, --help            show this help message and exit
  -i, --inline          Replace the file inline instead of returning the output to stdout
  -o TARGET, --output TARGET
                        If specified, the output will be written to TARGET filename instead of stdout
  --ensure_ascii        Pass ensure_ascii=True to json.dumps()
  --indent INDENT       Number of spaces for indentation (Default 2)

Please pin this library only on the major version!

We use TDD and strict semantic versioning, there will be frequent updates and no breaking changes in minor and patch versions. To ensure that you only pin the major version of this library in your requirements.txt, specify the package name followed by the major version and a wildcard for minor and patch versions. For example:

json_repair==0.*

In this example, any version that starts with 0. will be acceptable, allowing for updates on minor and patch versions.

If you are using this library in your academic work (as I know many folks are) please find the BibTex here:

@software{Baccianella_JSON_Repair_-_2024,
    author = {Baccianella, Stefano},
    month = aug,
    title = {{JSON Repair - A python module to repair invalid JSON, commonly used to parse the output of LLMs}},
    url = {https://github.com/mangiucugna/json_repair},
    version = {0.28.3},
    year = {2024}
}

Thank you for citing my work and please send me a link to the paper if you can!

This module will parse the JSON file following the BNF definition:

<json> ::= <primitive> | <container>

<primitive> ::= <number> | <string> | <boolean>
; Where:
; <number> is a valid real number expressed in one of a number of given formats
; <string> is a string of valid characters enclosed in quotes
; <boolean> is one of the literal strings 'true', 'false', or 'null' (unquoted)

<container> ::= <object> | <array>
<array> ::= '[' [ <json> *(', ' <json>) ] ']' ; A sequence of JSON values separated by commas
<object> ::= '{' [ <member> *(', ' <member>) ] '}' ; A sequence of 'members'
<member> ::= <string> ': ' <json> ; A pair consisting of a name, and a JSON value

If something is wrong (a missing parentheses or quotes for example) it will use a few simple heuristics to fix the JSON string:

• Add the missing parentheses if the parser believes that the array or object should be closed
• Quote strings or add missing single quotes
• Adjust whitespaces and remove line breaks

I am sure some corner cases will be missing, if you have examples please open an issue or even better push a PR

Just create a virtual environment with requirements.txt, the setup uses pre-commit to make sure all tests are run.

Make sure that the Github Actions running after pushing a new commit don't fail as well.

You will need owner access to this repository

• Edit pyproject.toml and update the version number appropriately using semver notation
• Commit and push all changes to the repository before continuing or the next steps will fail
• Run python -m build
• Create a new release in Github, making sure to tag all the issues solved and contributors. Create the new tag, same as the one in the build configuration
• Once the release is created, a new Github Actions workflow will start to publish on Pypi, make sure it didn't fail

• Typescript: https://github.com/josdejong/jsonrepair
• Go: https://github.com/RealAlexandreAI/json-repair
• Ruby: https://github.com/sashazykov/json-repair-rb