partial-json-parser 0.2.0.1

Parse partial JSON generated by LLM

Partial JSON Parser

Sometimes we need LLM (Large Language Models) to produce structural information instead of natural language. The easiest way is to use JSON.

But before receiving the last token of response, the JSON is broken, which means you can't use JSON.parse to decode it. But we still want to stream the data to the user.

Here comes partial-json-parser, a lightweight and customizable library for parsing partial JSON strings. Here is a demo.

(Note that there is a JavaScript implementation too)

Installation

pip install partial-json-parser # or poetry / pdm / conda

partial-json-parser is implemented purely in Python, with good type hints.

Usage

Importing the library

You can import the loads function and the Allow object from the library like this:

from partial_json_parser import loads, Allow

The Allow object is just an Enum for options. It determines what types can be partial. types not included in allow only appears after its completion can be ensured.

Parsing complete / partial JSON strings

The loads function works just like the built-in json.loads when parsing a complete JSON string:

result = loads('{"key":"value"}')
print(result)  # Outputs: {'key': 'value'}

You can parse a partial JSON string by passing an additional parameter to the loads function. This parameter is a bitwise OR of the constants from the Allow flag:

(Note that you can directly import the constants you need from partial-json-parser.options)

from partial_json_parser import loads, Allow
from partial_json_parser.options import STR, OBJ

result = loads('{"key": "v', STR | OBJ)
print(result)  # Outputs: {'key': 'v'}

In this example, Allow.STR tells the parser that it's okay if a string is incomplete, and Allow.OBJ tells the parser so as a dict. The parser then try to return as much data as it can.

If you don't allow partial strings, then it will not add "key" to the object because "v is not close:

result = loads('{"key": "v', OBJ)
print(result)  # Outputs: {}

result = loads('{"key": "value"', OBJ)
print(result)  # Outputs: {'key': 'value'}

Similarity, you can parse partial lists or even partial special values if you allow it:

(Note that allow defaults to Allow.ALL)

result = loads('[ {"key1": "value1", "key2": [ "value2')
print(result)  # Outputs: [{'key1': 'value1', 'key2': ['value2']}]

result = loads("-Inf")
print(result)  # Outputs: -inf

Handling malformed JSON

If the JSON string is malformed, the parse function will throw an error:

loads("wrong")  # MalformedJSON: Malformed node or string on line 1

API Reference

loads(json_string, [allow_partial])

• json_string <string>: The JSON string to parse.
• allow_partial <Allow | int>: Specify what kind of partialness is allowed during JSON parsing (default: Allow.ALL).

Returns the parsed Python value.

Allow

Enum class that specifies what kind of partialness is allowed during JSON parsing. It has the following members:

• STR: Allow partial string.
• NUM: Allow partial number.
• ARR: Allow partial array.
• OBJ: Allow partial object.
• NULL: Allow partial null.
• BOOL: Allow partial boolean.
• NAN: Allow partial NaN.
• INFINITY: Allow partial Infinity.
• _INFINITY: Allow partial -Infinity.
• INF: Allow both partial Infinity and -Infinity.
• SPECIAL: Allow all special values.
• ATOM: Allow all atomic values.
• COLLECTION: Allow all collection values.
• ALL: Allow all values.

Testing

To run the tests for this library, you should clone the repository and install the dependencies:

git clone https://github.com/promplate/partial-json-parser.git
cd partial-json-parser
pdm install

Then, you can run the tests using Hypothesis and Pytest:

pdm test

Please note that while we strive to cover as many edge cases as possible, it's always possible that some cases might not be covered.

License

This project is licensed under the MIT License.