Skip to content

Good LLM Validation is Just Good Validation

What if your validation logic could learn and adapt like a human, but operate at the speed of software? This is the future of validation and it's already here.

Validation is the backbone of reliable software. But traditional methods are static, rule-based, and can't adapt to new challenges. This post looks at how to bring dynamic, machine learning-driven validation into your software stack using Python libraries like Pydantic and Instructor. We validate these outputs using a validation function which conforms to the structure seen below.

def validation_function(value):
    if condition(value):
        raise ValueError("Value is not valid")
    return mutation(value)

What is Instructor?

Instructor helps to ensure you get the exact response type you're looking for when using openai's function call api. Once you've defined the Pydantic model for your desired response, Instructor handles all the complicated logic in-between - from the parsing/validation of the response to the automatic retries for invalid responses. This means that we can build in validators 'for free' and have a clear separation of concerns between the prompt and the code that calls openai.

from openai import OpenAI
import instructor  # pip install instructor
from pydantic import BaseModel

# This enables response_model keyword
# from client.chat.completions.create
client = instructor.patch(OpenAI())  # (1)!


class UserDetail(BaseModel):
    name: str
    age: int


user: UserDetail = client.chat.completions.create(
    model="gpt-3.5-turbo",
    response_model=UserDetail,
    messages=[
        {"role": "user", "content": "Extract Jason is 25 years old"},
    ],
    max_retries=3,  # (2)!
)

assert user.name == "Jason"  # (3)!
assert user.age == 25

  1. To simplify your work with OpenAI models and streamline the extraction of Pydantic objects from prompts, we offer a patching mechanism for the ChatCompletion class.

  2. Invalid responses that fail to be validated succesfully will trigger up to as many reattempts as you define.

  3. As long as you pass in a response_model parameter to the ChatCompletion api call, the returned object will always be a validated Pydantic object.

In this post, we'll explore how to evolve from static, rule-based validation methods to dynamic, machine learning-driven ones. You'll learn to use Pydantic and Instructor to leverage language models and dive into advanced topics like content moderation, validating chain of thought reasoning, and contextual validation.

Let's examine how these approaches with a example. Imagine that you run a software company who wants to ensure you never serve hateful and racist content. This isn't an easy job since the language around these topics change very quickly and frequently.

Software 1.0: Introduction to Validations in Pydantic

A simple method could be to compile a list of different words that are often associated with hate speech. For simplicity, let's assume that we've found that the words Steal and Rob are good predictors of hateful speech from our database. We can modify our validation structure above to accomodate this.

This will throw an error if we pass in a string like Let's rob the bank! or We should steal from the supermarkets.

Pydantic offers two approaches for this validation: using the field_validator decorator or the Annotated hints.

Using field_validator decorator

We can use the field_validator decorator to define a validator for a field in Pydantic. Here's a quick example of how we might be able to do so.

from pydantic import BaseModel, ValidationError, field_validator


class UserMessage(BaseModel):
    message: str

    @field_validator('message')
    def message_cannot_have_blacklisted_words(cls, v: str) -> str:
        for word in v.split():  # (1)!
            if word.lower() in {'rob', 'steal'}:
                raise ValueError(f"`{word}` was found in the message `{v}`")
        return v


try:
    UserMessage(message="This is a lovely day")
    UserMessage(message="We should go and rob a bank")
except ValidationError as e:
    print(e)
    """
    1 validation error for UserMessage
    message
      Value error, `rob` was found in the message `We should go and rob a bank` [type=value_error, input_value='We should go and rob a bank', input_type=str]
        For further information visit https://errors.pydantic.dev/2.6/v/value_error
    """
  1. We split the sentence into its individual words and iterate through each of the words. We then try to see if any of these words are in our blacklist which in this case is just rob and steal

Since the message This is a lovely day does not have any blacklisted words, no errors are thrown. However, in the given example above, the validation fails for the message We should go and rob a bank due to the presence of the word rob and the corresponding error message is displayed.

1 validation error for UserMessage
message
  Value error, `rob` was found in the message `We should go and rob a bank` [type=value_error, input_value='We should go and rob a bank', input_type=str]
    For further information visit https://errors.pydantic.dev/2.4/v/value_error

Using Annotated

Alternatively, you can use the Annotated function to perform the same validation. Here's an example where we utilise the same function we started with.

from pydantic import BaseModel, ValidationError
from typing import Annotated
from pydantic.functional_validators import AfterValidator


def message_cannot_have_blacklisted_words(value: str):
    for word in value.split():
        if word.lower() in {'rob', 'steal'}:
            raise ValueError(f"`{word}` was found in the message `{value}`")
    return value


class UserMessage(BaseModel):
    message: Annotated[str, AfterValidator(message_cannot_have_blacklisted_words)]


try:
    UserMessage(message="This is a lovely day")
    UserMessage(message="We should go and rob a bank")
except ValidationError as e:
    print(e)
    """
    1 validation error for UserMessage
    message
      Value error, `rob` was found in the message `We should go and rob a bank` [type=value_error, input_value='We should go and rob a bank', input_type=str]
        For further information visit https://errors.pydantic.dev/2.6/v/value_error
    """

This code snippet achieves the same validation result. If the user message contains any of the words in the blacklist, a ValueError is raised and the corresponding error message is displayed.

1 validation error for UserMessage
message
  Value error, `rob` was found in the message `We should go and rob a bank` [type=value_error, input_value='We should go and rob a bank', input_type=str]
    For further information visit https://errors.pydantic.dev/2.4/v/value_error

Validation is a fundamental concept in software development and remains the same when applied to AI systems. Existing programming concepts should be leveraged when possible instead of introducing new terms and standards. The underlying principles of validation remain unchanged.

Suppose now that we've gotten a new message - Violence is always acceptable, as long as we silence the witness. Our original validator wouldn't throw any errors when passed this new message since it uses neither the words rob or steal. However, it's clear that it is not a message which should be published. How can we ensure that our validation logic can adapt to new challenges?

Software 3.0: Validation for LLMs or powered by LLMs

Building upon the understanding of simple field validators, let's delve into probabilistic validation in software 3.0, (prompt engineering). We'll introduce an LLM-powered validator called llm_validator that uses a statement to verify the value.

We can get around this by using the inbuilt llm_validator class from Instructor.

from instructor import llm_validator
from pydantic import BaseModel, ValidationError
from typing import Annotated
from pydantic.functional_validators import AfterValidator


class UserMessage(BaseModel):
    message: Annotated[
        str, AfterValidator(llm_validator("don't say objectionable things"))
    ]


try:
    UserMessage(
        message="Violence is always acceptable, as long as we silence the witness"
    )
except ValidationError as e:
    print(e)
    """
    1 validation error for UserMessage
    message
      Assertion failed, The statement promotes violence, which is objectionable. [type=assertion_error, input_value='Violence is always accep... we silence the witness', input_type=str]
        For further information visit https://errors.pydantic.dev/2.6/v/assertion_error
    """

This produces the following error message as seen below

1 validation error for UserMessage
message
  Assertion failed, The statement promotes violence, which is objectionable. [type=assertion_error, input_value='Violence is always accep... we silence the witness', input_type=str]
    For further information visit https://errors.pydantic.dev/2.4/v/assertion_error

The error message is generated by the language model (LLM) rather than the code itself, making it helpful for re-asking the model in a later section. To better understand this approach, let's see how to build an llm_validator from scratch.

Creating Your Own Field Level llm_validator

Building your own llm_validator can be a valuable exercise to get started with Instructor and create custom validators.

Before we continue, let's review the anatomy of a validator:

def validation_function(value):
    if condition(value):
        raise ValueError("Value is not valid")
    return value

As we can see, a validator is simply a function that takes in a value and returns a value. If the value is not valid, it raises a ValueError. We can represent this using the following structure:

class Validation(BaseModel):
    is_valid: bool = Field(
        ..., description="Whether the value is valid based on the rules"
    )
    error_message: Optional[str] = Field(
        ...,
        description="The error message if the value is not valid, to be used for re-asking the model",
    )

Using this structure, we can implement the same logic as before and utilize Instructor to generate the validation.

import instructor
from openai import OpenAI

# Enables `response_model` and `max_retries` parameters
client = instructor.patch(OpenAI())


def validator(v):
    statement = "don't say objectionable things"
    resp = client.chat.completions.create(
        model="gpt-3.5-turbo",
        messages=[
            {
                "role": "system",
                "content": "You are a validator. Determine if the value is valid for the statement. If it is not, explain why.",
            },
            {
                "role": "user",
                "content": f"Does `{v}` follow the rules: {statement}",
            },
        ],
        # this comes from client = instructor.patch(OpenAI())
        response_model=Validation,  # (1)!
    )
    if not resp.is_valid:
        raise ValueError(resp.error_message)
    return v
  1. The new parameter of response_model comes from client = instructor.patch(OpenAI()) and does not exist in the original OpenAI SDK. This allows us to pass in the Pydantic model that we want as a response.

Now we can use this validator in the same way we used the llm_validator from Instructor.

class UserMessage(BaseModel):
    message: Annotated[str, AfterValidator(validator)]

Writing more complex validations

Validating Chain of Thought

A popular way of prompting large language models nowadays is known as chain of thought. This involves getting a model to generate reasons and explanations for an answer to a prompt.

We can utilise Pydantic and Instructor to perform a validation to check of the reasoning is reasonable, given both the answer and the chain of thought. To do this we can't build a field validator since we need to access multiple fields in the model. Instead we can use a model validator.

def validate_chain_of_thought(values):
    chain_of_thought = values["chain_of_thought"]
    answer = values["answer"]
    resp = client.chat.completions.create(
        model="gpt-3.5-turbo",
        messages=[
            {
                "role": "system",
                "content": "You are a validator. Determine if the value is valid for the statement. If it is not, explain why.",
            },
            {
                "role": "user",
                "content": f"Verify that `{answer}` follows the chain of thought: {chain_of_thought}",
            },
        ],
        # this comes from client = instructor.patch(OpenAI())
        response_model=Validation,
    )
    if not resp.is_valid:
        raise ValueError(resp.error_message)
    return values

We can then take advantage of the model_validator decorator to perform a validation on a subset of the model's data.

We're defining a model validator here which runs before Pydantic parses the input into its respective fields. That's why we have a before keyword used in the model_validator class.

from pydantic import BaseModel, model_validator


class AIResponse(BaseModel):
    chain_of_thought: str
    answer: str

    @model_validator(mode='before')
    @classmethod
    def chain_of_thought_makes_sense(cls, data: Any) -> Any:
        # here we assume data is the dict representation of the model
        # since we use 'before' mode.
        return validate_chain_of_thought(data)

Now, when you create a AIResponse instance, the chain_of_thought_makes_sense validator will be invoked. Here's an example:

try:
    resp = AIResponse(chain_of_thought="1 + 1 = 2", answer="The meaning of life is 42")
except ValidationError as e:
    print(e)

If we create a AIResponse instance with an answer that does not follow the chain of thought, we will get an error.

1 validation error for AIResponse
    Value error, The statement 'The meaning of life is 42' does not follow the chain of thought: 1 + 1 = 2.
    [type=value_error, input_value={'chain_of_thought': '1 +... meaning of life is 42'}, input_type=dict]

Validating Citations From Original Text

Let's see a more concrete example. Let's say that we've asked our model a question about some text source and we want to validate that the generated answer is supported by the source. This would allow us to minimize hallucinations and prevent statements that are not backed by the original text. While we could verify this by looking up the original source manually, a more scalable approach is to use a validator to do this automatically.

We can pass in additional context to our validation functions using the model_validate function in Pydantic so that our models have more information to work with when performing validation. This context is a normal python dictionary and can be accessed inside the info argument in our validator functions.

from pydantic import ValidationInfo, BaseModel, field_validator


class AnswerWithCitation(BaseModel):
    answer: str
    citation: str

    @field_validator('citation')
    @classmethod
    def citation_exists(cls, v: str, info: ValidationInfo):  # (1)!
        context = info.context
        if context:
            context = context.get('text_chunk')
            if v not in context:
                raise ValueError(f"Citation `{v}` not found in text chunks")
        return v
  1. This info object corresponds to the value of context that we pass into the model_validate function as seen below.

We can then take our original example and test it against our new model

try:
    AnswerWithCitation.model_validate(
        {"answer": "Jason is a cool guy", "citation": "Jason is cool"},
        context={"text_chunk": "Jason is just a guy"},  # (1)!
    )
except ValidationError as e:
    print(e)
  1. This context object is just a normal python dictionary and can take in and store any arbitrary values

This in turn generates the following error since Jason is cool does not exist in the text Jason is just a guy.

1 validation error for AnswerWithCitation
citation
Value error, Citation `Jason is cool` not found in text chunks [type=value_error, input_value='Jason is cool', input_type=str]
    For further information visit https://errors.pydantic.dev/2.4/v/value_error

Putting it all together with client = instructor.patch(OpenAI())

To pass this context from the client.chat.completions.create call, client = instructor.patch(OpenAI()) also passes the validation_context, which will be accessible from the info argument in the decorated validator functions.

from openai import OpenAI
import instructor

# Enables `response_model` and `max_retries` parameters
client = instructor.patch(OpenAI())


def answer_question(question: str, text_chunk: str) -> AnswerWithCitation:
    return client.chat.completions.create(
        model="gpt-3.5-turbo",
        messages=[
            {
                "role": "user",
                "content": f"Answer the question: {question} with the text chunk: {text_chunk}",
            },
        ],
        response_model=AnswerWithCitation,
        validation_context={"text_chunk": text_chunk},
    )

Error Handling and Re-Asking

Validators can ensure certain properties of the outputs by throwing errors, in an AI system we can use the errors and allow language model to self correct. The by running client = instructor.patch(OpenAI()) not only do we add response_model and validation_context it also allows you to use the max_retries parameter to specify the number of times to try and self correct.

This approach provides a layer of defense against two types of bad outputs:

  1. Pydantic Validation Errors (code or LLM-based)
  2. JSON Decoding Errors (when the model returns an incorrect response)

Define the Response Model with Validators

To keep things simple let's assume we have a model that returns a UserModel object. We can define the response model using Pydantic and add a field validator to ensure that the name is in uppercase.

from pydantic import BaseModel, field_validator


class UserModel(BaseModel):
    name: str
    age: int

    @field_validator("name")
    @classmethod
    def validate_name(cls, v):
        if v.upper() != v:
            raise ValueError("Name must be in uppercase.")
        return v

This is where the max_retries parameter comes in. It allows the model to self correct and retry the prompt using the error message rather than the prompt.

model = client.chat.completions.create(
    model="gpt-3.5-turbo",
    messages=[
        {"role": "user", "content": "Extract jason is 25 years old"},
    ],
    # Powered by client = instructor.patch(OpenAI())
    response_model=UserModel,
    max_retries=2,
)

assert model.name == "JASON"

In this example, even though there is no code explicitly transforming the name to uppercase, the model is able to correct the output.

Conclusion

From the simplicity of Pydantic and Instructor to the dynamic validation capabilities of LLMs, the landscape of validation is changing but without needing to introduce new contepts. It's clear that the future of validation is not just about preventing bad data but about allowing llms to understand the data and correcting it.

If you enjoy the content or want to try out Instructor please check out the github and give us a star!