instructor/docs/reask_validation.md

Integrated Validation and Reask with LLMs and Pydantic

Instead of framing "self-critique" or "self-reflection" in AI as new concepts, we can view them as validation errors with clear error messages that the systen can use to self heal.

Applications and Scenarios

• Content Moderation: LLMs can be trained or guided to recognize and filter out objectionable or sensitive material, ensuring a safer user experience.
• Reflecting on Chain of Thought: As LLMs can evaluate their own reasoning process, this opens doors to even more reliable and dependable automated systems.
• Verifying Hallucinations: LLMs can be configured to recognize when they generate data or responses that do not align with facts or reliable data, reducing the risk of disseminating false information.
• Data Integrity: Enforces data quality standards.

Pythonic Validation with Pydantic and Instructor

1. Uniform Validation API: Pydantic provides identical developer experience, whether using code-based or LLM-based validation.
2. Reasking Mechanism: Pydantic accumulates validation errors for a one-step reasking process.
3. Prompt Chaining via Error Messages: Instructor utilizes validation error messages to refine LLM outputs without and new abstractions.

Uniform Validation: Code-Based vs. LLM

Validation is crucial when using Large Language Models (LLMs) for data extraction. It ensures data integrity, ensuring both quantitative and qualititave correctness with code and llm validations.

!!! note "Pydantic Validation Docs" Pydantic supports validation individual fields or the whole model dict all at once.

- [Field-Level Validation](https://docs.pydantic.dev/latest/usage/validators/)
- [Model-Level Validation](https://docs.pydantic.dev/latest/usage/validators/#model-validators)

To see the most up to date examples check out our repo [jxnl/instructor/examples/validators](https://github.com/jxnl/instructor/tree/main/examples/validators)

Code-Based Validation Example

!!! note "Model Level Evaluation" Right now we only go over the field level examples, check out Model-Level Validation if you want to see how to do model level evaluation

Enforce a naming rule using Pydantic's built-in validation:

from pydantic import BaseModel, ValidationError
from typing_extensions import Annotated
from pydantic import AfterValidator

def name_must_contain_space(v: str) -> str:
    if " " not in v:
        raise ValueError("Name must contain a space.")
    return v.lower()

class UserDetail(BaseModel):
    age: int
    name: Annotated[str, AfterValidator(name_must_contain_space)]

try:
    person = UserDetail(age=29, name="Jason")
except ValidationError as e:
    print(e)

Output for Code-Based Validation

1 validation error for UserDetail
name
   Value error, name must contain a space (type=value_error)

LLM-Based Validation Example

LLM-based validation can also be plugged into the same Pydantic model. Here, if the answer attribute contains content that violates the rule "don't say objectionable things," Pydantic will raise a validation error.

from pydantic import BaseModel, ValidationError, BeforeValidator
from typing_extensions import Annotated
from instruct import llm_validator

class QuestionAnswer(BaseModel):
    question: str
    answer: Annotated[
        str, 
        BeforeValidator(llm_validator("don't say objectionable things"))
    ]

try:
    qa = QuestionAnswer(
        question="What is the meaning of life?",
        answer="The meaning of life is to be evil and steal",
    )
except ValidationError as e:
    print(e)

Output for LLM-Based Validation

Its important to not here that the error message is generated by the LLM, not the code, so it'll be helpful for re asking the model.

1 validation error for QuestionAnswer
answer
   Assertion failed, The statement is objectionable. (type=assertion_error)

Using Reasking Logic to Correct Outputs

Validators are a great tool for ensuring some property of the outputs. When you use the patch() method with the openai client, you can use the max_retries parameter to set the number of times you can reask the model to correct the output.

Its a great layer of defense against bad outputs of two forms.

1. Pydantic Validation Errors (code or llm based)
2. JSON Decoding Errors (when the model returns a bad response)

Step 1: Define the Response Model with Validators

Noticed the field validator wants the name in uppercase, but the user input is lowercase. The validator will raise a ValueError if the name is not in uppercase.

import instructor
from pydantic import BaseModel, field_validator

# Apply the patch to the OpenAI client
instructor.patch()

class UserDetails(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

Step 2. Using the Client with Retries

Here, the UserDetails model is passed as the response_model, and max_retries is set to 2.

model = openai.ChatCompletion.create(
    model="gpt-3.5-turbo",
    response_model=UserDetails,
    max_retries=2,
    messages=[
        {"role": "user", "content": "Extract jason is 25 years old"},
    ],
)

assert model.name == "JASON"

What happens behind the scenes?

Behind the scenes, the instructor.patch() method adds a max_retries parameter to the openai.ChatCompletion.create() method. The max_retries parameter will trigger up to 2 reattempts if the name attribute fails the uppercase validation in UserDetails.

try:
    ...
except (ValidationError, JSONDecodeError) as e:
    kwargs["messages"].append(dict(**response.choices[0].message))
    kwargs["messages"].append(
        {
            "role": "user",
            "content": f"Please correct the function call; errors encountered:\n{e}",
        }
    )

Advanced Validation Techniques

The docs are currently incomplete, but we have a few advanced validation techniques that we're working on documenting better, for a example of model level validation, and using a validation context check out our example on verifying citations which covers

1. Validate the entire object with all attributes rather than one attribute at a time
2. Using some 'context' to validate the object, in this case we use the context to check if the citation existed in the original text.

Takeaways

By integrating these advanced validation techniques, we not only improve the quality and reliability of LLM-generated content but also pave the way for more autonomous and effective systems.