Skip to main content

Documentation Index

Fetch the complete documentation index at: https://snowglobe.so/docs/llms.txt

Use this file to discover all available pages before exploring further.

If you need to perform a validation that is not currently supported by the hub, you can create your own custom validators.

​
As a function

A custom validator can be as simple as a single function if you do not require additional arguments: 
from typing import Dict
from guardrails.validators import (
    FailResult,
    PassResult,
    register_validator,
    ValidationResult,
)

@register_validator(name="toxic-words", data_type="string")
def toxic_words(value, metadata: Dict) -> ValidationResult:
    mentioned_words = []
    for word in ["butt", "poop", "booger"]:
        if word in value:
            mentioned_words.append(word)

    if len(mentioned_words) > 0:
        return FailResult(
            error_message=f"Mention toxic words: {', '.join(mentioned_words)}",
        )
    else:
        return PassResult()

​
As a class

If you need to perform more complex operations or require additional arguments to perform the validation, then the validator can be specified as a class that inherits from our base Validator class: 
from typing import Callable, Dict, Optional, List
from guardrails.validators import (
    FailResult,
    PassResult,
    register_validator,
    ValidationResult,
    Validator,
)

@register_validator(name="toxic-words", data_type="string")
class ToxicWords(Validator):
    def __init__(self, search_words: List[str]=["booger", "butt"], on_fail: Optional[Callable] = None):
        super().__init__(on_fail=on_fail, search_words=search_words)
        self.search_words = search_words

    def _validate(self, value: List[str], metadata: Dict) -> ValidationResult:
        mentioned_words = []
        for word in self.search_words:
            if word in value:
                mentioned_words.append(word)

        if len(mentioned_words) > 0:
            return FailResult(
                error_message=f"Mentioned toxic words: {', '.join(mentioned_words)}",
            )
        else:
            return PassResult()

​
On fail

In the code below, a fix_value will be supplied in the FailResult. This value will represent a programmatic fix that can be applied to the output if on_fail='fix' is passed during validator initialization.
For more details about on fail actions refer to: On fail actions
from typing import Callable, Dict, Optional
from guardrails.validators import (
    FailResult,
    PassResult,
    register_validator,
    ValidationResult,
    Validator,
)

@register_validator(name="toxic-words", data_type="string")
class ToxicWords(Validator):
    def __init__(self, search_words: List[str]=["booger", "butt"], on_fail: Optional[Callable] = None):
        super().__init__(on_fail=on_fail, search_words=search_words)
        self.search_words = search_words

    def _validate(self, value: str, metadata: Dict) -> ValidationResult:
        mentioned_words = []
        for word in self.search_words:
            if word in value:
                mentioned_words.append(word)

        if len(mentioned_words) > 0:
            # Filter out the toxic words from the value
            on_fix = ' '.join([word for word in value.split() if word not in self.search_words])
            return FailResult(
                error_message=f"Value {value} does mention words: {', '.join(mentioned_words)}",
                on_fix=on_fix,
            )
        else:
            return PassResult()

​
Model and LLM integration

Validators are not limited to just programmatic and algorithmic validation. In this example we integrate with Hugging Face to validate if a value meets a threshold for toxic language. 
from typing import Callable, Dict, Optional
from guardrails.validators import (
    FailResult,
    PassResult,
    register_validator,
    ValidationResult,
    Validator,
)
from litellm import completion
from transformers import pipeline

@register_validator(name="toxic-language", data_type="string")
class ToxicLanguageValidator(Validator):
    def __init__(
            self, 
            threshold: float = 0.9,
            device: int = -1,  # Add device parameter with default value -1 (CPU). 0 is GPU.
            model_name: str = "unitary/toxic-bert",
            on_fail: Optional[Callable] = None
            ):
        super().__init__(on_fail=on_fail, threshold=threshold)
        self._threshold = threshold
        self.pipeline = pipeline("text-classification", model=model_name, device=device)

    def _validate(self, value: str, metadata: Dict) -> ValidationResult:
        result = self.pipeline(value)
        if result[0]['label'] == 'toxic' and result[0]['score'] > self._threshold:
            return FailResult(
                error_message=f"{value} failed validation. Detected toxic language with score {result[0]['score']}."
            )
        else:
            return PassResult()
The validator below can similarly identify toxic language but instead of using a model trained for it specifically it uses chat-gpt. 
from typing import Callable, Dict, Optional
from guardrails.validators import (
    FailResult,
    PassResult,
    register_validator,
    ValidationResult,
    Validator,
)
from litellm import completion

PROMPT = """
You are a helpful assistant.
Analyze the other messages provided and respond with a number between 0 and 100 and nothing else.
Do not output prose.
0 means the messages contain no toxic language.
100 means the messages contain very toxic language that is offensive, vulgar, negative or insulting.
"""

@register_validator(name="toxic-language", data_type="string")
class ToxicLanguage(Validator):
    def __init__(self, threshold: int=70, model='gpt-4o', on_fail: Optional[Callable] = None):
        super().__init__(on_fail=on_fail, threshold=threshold)
        self._threshold = threshold
        self._model = model

    def _llm_callable(self, messages):
        return completion(
            model=self._model,
            messages=messages,
            temperature=0.0,
        )
    
    def _validate(self, value: str, metadata: Dict) -> ValidationResult:
        messages = [
            {
                "role": "system",
                "content": PROMPT,
            },
            {
                "role": "user",
                "content": value,
            }
        ]

        score = int(self._llm_callable(messages).choices[0].message.content)
        if score > self._threshold:
            return FailResult(
                error_message=f"{value} failed validation. Score {score} is below threshold of {self._threshold}."
            )
        else:
            return PassResult()

​
Streaming

Validators support streaming validation out of the box. The validate_stream method handles calling _validate with accumulated chunks of a stream when a guard is executed with guard(streaming=True, ...). By default stream validation is done on a per sentence basis. Validator._chunking_function may be overloaded to provide a custom chunking strategy. This is useful to optimize latency when integrating outside services such as LLMs and controlling how much data a validating model gets to give it more or less context. The code below in a validator will cause a validator to validate a stream of text 1 paragraph at a time. 
@register_validator(name="toxic-language", data_type="string")
class ToxicLanguageValidator(Validator):
    def __init__(
            self, 
            threshold: float = 0.9,
            device: int = -1,  # Add device parameter with default value -1 (CPU). 0 is GPU.
            model_name: str = "unitary/toxic-bert",
            on_fail: Optional[Callable] = None
            ):
        super().__init__(on_fail=on_fail, threshold=threshold)
        self._threshold = threshold
        self.pipeline = pipeline("text-classification", model=model_name, device=device)

    def _validate(self, value: str, metadata: Dict) -> ValidationResult:
        result = self.pipeline(value)
        if result[0]['label'] == 'toxic' and result[0]['score'] > self._threshold:
            return FailResult(
                error_message=f"{value} failed validation. Detected toxic language with score {result[0]['score']}."
            )
        else:
            return PassResult()

    def _chunking_function(chunk: str) -> List[str]:
        """The strategy used for chunking accumulated text input into
        validation sets of a paragraph

        Args:
            chunk (str): The text to chunk into some subset.

        Returns:
            list[str]: The text chunked into some subset.
        """
        if "\n\n" not in chunk:
            return []
        fragments = chunk.split("\n\n")
        return [fragments[0] + "\n\n", "\n\n".join(fragments[1:])]

​
Usage

Custom validators must be defined before creating a Guard or RAIL spec in the code, but otherwise can be used like built in validators.

​
Guard.use example

from guardrails import Guard
from .my_custom_validators import toxic_words, ToxicLanguage

guard = Guard().use(
    ToxicLanguage(threshold=0.8)
).use(
    toxic_words()
)

​
Pydantic example

from guardrails import Guard
from pydantic import BaseModel, Field
from .my_custom_validators import toxic_words, ToxicLanguage

class MyModel(BaseModel):
    a_string: Field(validators=[toxic_words()])
    custom_string: Field(validators=[ToxicLanguage(threshold=0.8)])

guard = Guard.for_pydantic(MyModel)

​
RAIL example

<rail version="0.1">
...
<output>
    <string name="a_string" type="string" validators="toxic-words" on-fail-toxic-words="exception" />
    <string name="custom_string" type="string" validators="toxic-language:threshold=0.4" />
</output>
...
</rail>