Integrating Instructor with SQLModel¶

SQLModel is a library designed for interacting with SQL databases from Python code using Python objects.

SQLModel is based on Pydantic and SQLAlchemy and was created by tiangolo who also developed FastAPI.

So you can expect seamless integration across all these libraries, reducing code duplicating and improving your developer experience.

Example: Adding responses from Instructor directly to your DB¶

Defining the Models¶

First we'll define a model that will serve as a table for our database and the structure of our outputs from Instructor

from typing import Optional
from uuid import UUID, uuid4
from pydantic.json_schema import SkipJsonSchema
from sqlmodel import Field, SQLModel
import instructor


class Hero(SQLModel, instructor.OpenAISchema, table=True):
    id: SkipJsonSchema[UUID] = Field(default_factory=lambda: uuid4(), primary_key=True)
    name: str
    secret_name: str
    age: Optional[int] = None

The Importance of Using SkipJsonSchema¶

Notice the use of SkipJsonSchema for the id field.

This prunes the field from the JSON schema sent to the LLM, so it won't try to generate a UUID.

When Instructor unpacks the response and loads it into the Hero model, it will automatically generate a UUID using the default_factory.

This approach saves tokens during LLM generation and more importantly protects against errors that might occur if the LLM generates an incorrect UUID format.

The resulting JSON schema sent to the LLM will look like:

{
    "properties": {
        "name": {
            "title": "Name",
            "type": "string"
        },
        "secret_name": {
            "title": "Secret Name",
            "type": "string"
        },
        "age": {
            "anyOf": [
                {
                    "type": "integer"
                },
                {
                    "type": "null"
                }
            ],
            "default": null,
            "title": "Age"
        }
    },
    "required": [
        "name",
        "secret_name"
    ],
    "title": "Hero",
    "type": "object"
}

Generating a record¶

The create_hero function will query OpenAI for a Hero record

import instructor
from openai import OpenAI


client = instructor.from_openai(OpenAI())


def create_hero() -> Hero:
    return client.chat.completions.create(
        model="gpt-3.5-turbo",
        response_model=Hero,
        messages=[
            {"role": "user", "content": "Make a new superhero"},
        ],
    )

Inserting the response into the DB¶

engine = create_engine("sqlite:///database.db")
SQLModel.metadata.create_all(engine)

hero = create_hero()

# The Raw Response from the LLM will not have an id due to the SkipJsonSchema
print(hero._raw_response.choices[0].message.content)
#> {'name': 'Superman', 'secret_name': 'Clark Kent', 'age': 30}

# The model_dump() method will include the generated id as it has been loaded as a Hero object
print(hero.model_dump())
#> {'name': 'Superman', 'secret_name': 'Clark Kent', 'age': 30, 'id': UUID('1234-5678-...')}

with Session(engine) as session:
    session.add(hero)
    session.commit()

And there you have it! You can now use the same models for your database and Instructor enabling them work seamlessly! Also checkout the FastAPI guide to see how you can use these models in an API as well.