FuzzTypes 0.0.2

FuzzTypes is a Pydantic extension for annotating autocorrecting fields

FuzzTypes

FuzzTypes is a set of "autocorrecting" annotation types that expands upon Pydantic's included data conversions. Designed for simplicity, it provides powerful normalization capabilities (e.g. named entity linking) to ensure structured data is composed of "smart things" not "dumb strings".

Basic Use Case

Pydantic supports basic conversion of data between types. For instance:

from pydantic import BaseModel

class Normal(BaseModel):
    boolean: bool
    float: float
    integer: int
    
obj = Normal(
    boolean='yes',
    float='2',
    integer='3',
)
assert obj.boolean is True
assert obj.float == 2.0
assert obj.integer == 3

FuzzTypes expands on the standard data conversions handled by Pydantic and provides a variety of autocorrecting annotation types.

from pydantic import BaseModel
from fuzztypes import (
    ASCII,
    Datetime,
    Email,
    Fuzzmoji,
    InMemory,
    Integer,
    Person,
    Regex,
    ZipCode,
    flags,
)

inventors = ["Ada Lovelace", "Alan Turing", "Claude Shannon"]
Inventor = InMemory(inventors, search_flag=flags.FuzzSearch)
Handle =  Regex(r'@\w{1,15}', examples=["@genomoncology"])


class Fuzzy(BaseModel):
    ascii: ASCII
    email: Email
    emoji: Fuzzmoji
    handle: Handle
    integer: Integer
    inventor: Inventor 
    person: Person
    time: Datetime
    zipcode: ZipCode
    
obj = Fuzzy(
    ascii="άνθρωπος",
    email="John Doe <[e-mail]>",
    emoji='thought bubble',
    handle='Ian Maurer (@imaurer)',
    integer='fifty-five',
    inventor='ada luvlace',
    person='mr. arthur herbert fonzarelli (fonzie)',
    time='5am on Jan 1, 2025',
    zipcode="(Zipcode: 12345-6789)",
)
# greek for man: https://en.wiktionary.org/wiki/άνθρωπος
assert obj.ascii == "anthropos"
# extract email via regular expression
assert obj.email == "[e-mail]"
# fuzzy match "thought bubble" to "thought balloon" emoji
assert obj.emoji == "💭"
# simple, inline regex example (see above Handle type)
assert obj.handle == "@imaurer"
# convert integer word phrase to integer value
assert obj.integer == 55
# case-insensitive fuzzy match on lowercase, misspelled name
assert obj.inventor == "Ada Lovelace"
# human name parser (title, first, middle, last, suffix, nickname)
assert str(obj.person) == 'Mr. Arthur Herbert Fonzarelli (fonzie)'
assert obj.person.short_name == "Arthur Fonzarelli"
assert obj.person.nickname == "fonzie"
assert obj.person.last == "Fonzarelli"
# convert time phrase to datetime object
assert obj.time.isoformat() == "2025-01-01T05:00:00"
# extract zip5 or zip9 formats using regular expressions
assert obj.zipcode == "12345-6789"

Types can also be used outside of Pydantic models to validate and normalize data:

from fuzztypes import Date, Fuzzmoji

# access value via "call" (parenthesis)
assert Date("1 JAN 2023").isoformat() == "2023-01-01"
assert Fuzzmoji("tada") == '🎉'

# access entity via "key lookup" (square brackets)
assert Fuzzmoji["movie cam"].value == "🎥"
assert Fuzzmoji["movie cam"].aliases == [':movie_camera:', 'movie camera']
assert Fuzzmoji["movie cam"].model_dump() == {
    'value': '🎥',
    'label': None,
    'meta': None,
    'priority': None,
    'aliases': [':movie_camera:', 'movie camera']
}

Installation

Available on PyPI:

pip install fuzztypes

Structured Data Generation Use Case

Several libraries (e.g. Instructor, Outlines, Marvin) use Pydantic to define models for structured data generation using Large Language Models (LLMs) via function calling or a grammar/regex based sampling approach based on the JSON schema generated by Pydantic.

This approach allows for the enumeration of allowed values using Python's Literal, Enum or JSON Schema's examples field directly in your Pydantic class declaration which is used by the LLM to generate valid values. This approach works exceptionally well for low-cardinality (not many unique allowed values) such as the world's continents (7 in total).

This approach, however, doesn't scale well for high-cardinality (many unique allowed values) such as the number of known human genomic variants (~325M). Where exactly the cutoff is between "low" and "high" cardinality is an exercise left to the reader and their use case.

That's where FuzzTypes come in. The allowed values are managed by the FuzzTypes annotations and the values are resolved during the Pydantic validation process. This can include fuzzy and semantic searching that throws an exception if the provided value doesn't meet a minimum similarity threshold defined by the developer.

Errors discovered via Pydantic can be caught and resubmitted to the LLM for correction. The error will contain examples, expected patterns, and closest matches to help steer the LLM to provide a better informed guess.

Here are the updated "Base Types" and "Usable Types" sections with improved introductions and descriptions:

Base Types

Base types are the fundamental building blocks in FuzzTypes. They provide the core functionality and can be used to create custom annotation types tailored to specific use cases.

Type	Description
DateType	Base type for fuzzy parsing date objects.
Function	Allows using any function that accepts one value and returns one value for transformation.
InMemory	Enables matching entities in memory using exact, alias, fuzzy, or semantic search.
OnDisk	Performs matching entities stored on disk using exact, alias, fuzzy, or semantic search.
Regex	Allows matching values using a regular expression pattern.
TimeType	Base type for fuzzy parsing datetime objects (e.g., "tomorrow at 5am").

These base types offer flexibility and extensibility, enabling you to create custom annotation types that suit your specific data validation and normalization requirements.

Usable Types

Usable types are pre-built annotation types in FuzzTypes that can be directly used in Pydantic models. They provide convenient and ready-to-use functionality for common data types and scenarios.

Type	Description
ASCII	Converts Unicode strings to ASCII equivalents using either anyascii or unidecode.
Date	Converts date strings to date objects using dateparser.
Email	Extracts email addresses from strings using a regular expression.
Emoji	Matches emojis based on Unicode Consortium aliases using the emoji library.
Fuzzmoji	Matches emojis using fuzzy string matching against aliases.
Integer	Converts numeric strings or words to integers using number-parser.
Person	Parses person names into subfields (e.g., first, last, suffix) using python-nameparser.
SSN	Extracts U.S. Social Security Numbers from strings using a regular expression.
Time	Converts datetime strings to datetime objects using dateparser.
Vibemoji	Matches emojis using semantic similarity against aliases.
Zipcode	Extracts U.S. ZIP codes (5 or 9 digits) from strings using a regular expression.

These usable types provide a wide range of commonly needed data validations and transformations, making it easier to work with various data formats and perform tasks like parsing, extraction, and matching.

Configuring FuzzTypes

FuzzTypes provides a set of configuration options that allow you to customize the behavior of the annotation types. These options can be passed as arguments when creating an instance of a FuzzType.

The following table describes the available configuration options:

Argument	Type	Default	Description
case_sensitive	bool	False	If True, matches are case-sensitive. If False, matches are case-insensitive.
device	Literal["cpu", "cuda", "mps"]	"cpu"	The device to use for generating semantic embeddings and LanceDB indexing. Available options are "cpu", "cuda" (for NVIDIA GPUs), and "mps" (for Apple's Metal Performance Shaders).
encoder	Union[Callable, str, Any]	None	The encoder to use for generating semantic embeddings. It can be a callable function, a string specifying the name or path of a pre-trained model, or any other object that implements the encoding functionality.
examples	List[Any]	None	A list of example values to be used in schema generation. These examples are included in the generated JSON schema to provide guidance on the expected format of the input values.
fuzz_scorer	Literal["token_sort_ratio", ...]	"token_sort_ratio"	The scoring algorithm to use for fuzzy string matching. Available options include "token_sort_ratio", "ratio", "partial_ratio", "token_set_ratio", "partial_token_set_ratio", "token_ratio", "partial_token_ratio", "WRatio", and "QRatio". Each algorithm has its own characteristics and trade-offs between accuracy and performance.
limit	int	10	The maximum number of matches to return when performing fuzzy or semantic searches.
min_similarity	float	80.0	The minimum similarity score required for a match to be considered valid. Matches with a similarity score below this threshold will be discarded.
notfound_mode	Literal["raise", "none", "allow"]	"raise"	The action to take when a matching entity is not found. Available options are "raise" (raises an exception), "none" (returns None), and "allow" (returns the input key as the value).
search_flag	flags.SearchFlag	flags.DefaultSearch	The search strategy to use for finding matches. It is a combination of flags that determine which fields of the NamedEntity are considered for matching and whether fuzzy or semantic search is enabled. Available options are defined in the flags module.
tiebreaker_mode	Literal["raise", "lesser", "greater"]	"raise"	The strategy to use for resolving ties when multiple matches have the same similarity score. Available options are "raise" (raises an exception), "lesser" (returns the match with the lower value), and "greater" (returns the match with the greater value).
validator_mode	Literal["before"]	"before"	The validation mode to use for Pydantic. Currently, only the "before" mode is fully tested and supported, which resolves the value before validation.

These configuration options provide flexibility in tailoring the behavior of FuzzTypes to suit your specific use case. By adjusting these options, you can control aspects such as case sensitivity, device selection, encoding mechanism, search strategy, similarity thresholds, and more.

Lazy Dependencies

FuzzTypes leverages several powerful libraries to extend its functionality.

These dependencies are not installed by default with FuzzTypes to keep the installation lightweight. Instead, they are optional and can be installed as needed depending on which types you use.

Below is a list of these dependencies, including their licenses, purpose, and what specific Types require them.

Right now, you must pip install the modules directly, will be adding pip extr

Fuzz Type	Library	License	Purpose
ASCII	anyascii	ISC	Converting Unicode into ASCII equivalents (not GPL)
ASCII	unidecode	GPL	Converting Unicode into ASCII equivalents (better quality)
Date	dateparser	BSD-3	Parsing dates from strings
Emoji	emoji	BSD	Handling and manipulating emoji characters
Fuzz	rapidfuzz	MIT	Performing fuzzy string matching
InMemory	numpy	BSD	Numerical computing in Python
InMemory	scikit-learn	BSD	Machine learning in Python
InMemory	sentence-transformers	Apache-2.0	Encoding sentences into high-dimensional vectors
Integer	number-parser	BSD-3	Parsing numbers from strings
OnDisk	lancedb	Apache-2.0	High-performance, on-disk vector database
OnDisk	pyarrow	Apache-2.0	In-memory columnar data format and processing library
OnDisk	sentence-transformers	Apache-2.0	Encoding sentences into high-dimensional vectors
Person	nameparser	LGPL	Parsing person names

Maintainer

FuzzTypes was created by Ian Maurer, the CTO of GenomOncology.

This MIT-based open-source project was extracted from our product which includes the ability to normalize biomedical data for use in precision oncology clinical decision support systems. Contact me to learn more about our product offerings.

Roadmap

Additional capabilities will soon be added:

• Complete OnDisk fuzzy string matching.
• Reranking models
• Hybrid search (linear and reciprocal rank fusion using fuzzy and semantic)
• Trie-based autocomplete and aho-corasick search
• Humanize intword and ordinals
• Pint quantities
• Country and Currency codes/names

The following usable types are planned for future implementation in FuzzTypes:

Type	Description
AirportCode	Represents airport codes (e.g., "ORD").
Airport	Represents airport names (e.g., "O'Hare International Airport").
CountryCode	Represents ISO country codes (e.g., "US").
Country	Represents country names (e.g., "United States").
Currency	Represents currency codes (e.g., "USD").
LanguageCode	Represents ISO language codes (e.g., "en").
Language	Represents language names (e.g., "English").
Quantity	Converts strings to Quantity objects with value and unit using pint.
URL	Represents normalized URLs with tracking parameters removed using url-normalize.
USStateCode	Represents U.S. state codes (e.g., "CA").
USState	Represents U.S. state names (e.g., "California").