Skip to main content

structured llm outputs

Project description

struct-gpt

codecov PyPI

Get structured output from LLM's.

from struct_gpt import OpenAiBase
from pydantic import Field


class SentimentSchema(OpenAiBase):
    """
    Determine the sentiment of the given text:

    {content}
    """
    # 👆this becomes the prompt

    sentiment: str = Field(description="Either -1, 0, or 1.")


result = SentimentSchema.from_openai(content="I love pizza!").json()
                                     # 👆this ends up in the prompt
print(result)

outputs:

{
  "sentiment": "1"
}

Features

  • Easy creation of custom models using the OpenAI API
  • Integration with Pydantic for model validation and serialization
  • Flexible configuration with retries and temperature settings

Installation

pip install struct-gpt

Usage

  • Template variables in the class' docstring are replaced with the keyword arguments passed to from_openai
  • Classes can reference one another
  • You can use the OpenAiMixin to add functionality to existing Pydantic classes
from struct_gpt import OpenAiBase, OpenAiMixin
from pydantic import Field, BaseModel
from typing import Mapping


class SentimentSchema(OpenAiBase):
    """
    Determine the sentiment of the given text:

    {content}
    """

    sentiment: str = Field(description="Either -1, 0, or 1.")


class SentimentAnalysis(BaseModel, OpenAiMixin):
    """
    Determine the sentiment of each word in the following: {text}

    Also determine the overall sentiment of the text and who the narrator is.
    """

    words_to_sentiment: Mapping[str, SentimentSchema]
    overall_sentiment: SentimentSchema
    narrator: str


print(
    SentimentAnalysis.from_openai(
        text="As president, I loved the beautiful scenery, but the long hike was exhausting."
    ).json(indent=2)
)
See Output
{
  "words_to_sentiment": {
    "As": {
      "sentiment": "0"
    },
    "president,": {
      "sentiment": "1"
    },
    "I": {
      "sentiment": "0"
    },
    "loved": {
      "sentiment": "1"
    },
    "the": {
      "sentiment": "0"
    },
    "beautiful": {
      "sentiment": "1"
    },
    "scenery,": {
      "sentiment": "1"
    },
    "but": {
      "sentiment": "-1"
    },
    "long": {
      "sentiment": "-1"
    },
    "hike": {
      "sentiment": "-1"
    },
    "was": {
      "sentiment": "0"
    },
    "exhausting.": {
      "sentiment": "-1"
    }
  },
  "overall_sentiment": {
    "sentiment": "0"
  },
  "narrator": "president"
}

Improving reliability with examples

create can accept a list of examples to guide the model and improve its accuracy.

Each example is a dictionary containing an input and output key.

output should be an instance of the model, a dictionary, or its json string representation,.

from struct_gpt import OpenAiBase
from pydantic import Field


class SentimentSchema(OpenAiBase):
    """
    Determine the sentiment of the given text:

    {content}
    """

    sentiment: str = Field(description="Either -1, 0, or 1.")


examples = [
    {
        "input": "I love the beach!",
        "output": {"sentiment": "1"},
    },
    {
        "input": "don't touch that",
        "output": '{"sentiment": "-1"}',
    },
    {
        "input": "this library is neat!",
        "output": SentimentSchema(sentiment="1"),
    },
]

print(SentimentSchema.from_openai(content="I love pizza!", examples=examples).json())

outputs:

{
  "sentiment": "1"
}

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

struct_gpt-2.2.4.tar.gz (5.6 kB view hashes)

Uploaded Source

Built Distribution

struct_gpt-2.2.4-py3-none-any.whl (5.7 kB view hashes)

Uploaded Python 3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page