What is this?

Define a LLM based UDF for Apache Spark with simple code!

# Model for structured output
class Fruit(pydantic.BaseModel):
   name: str
   color: str
   taste: str

# Prompt with example
prompt = """
    return the color and taste of given fruit.

    #example

    ## input
    apple

    ## output
    {{
        "name": "apple",
        "color": "red",
        "taste": "sweet"
    }}
"""
# Simple UDF builder in openaivec
udf = UDFBuilder.of_openai(...)

# Register UDFs with structured output
spark.udf.register("parse_fruit", udf.completion(prompt, response_format=Fruit))

# Use UDFs in Spark SQL
spark.sql("SELECT name, parse_fruit(name) from dummy").show(truncate=False)

The following output is produced:

+------+--------------------------------------------------------+
|name  |fruit(name)                                             |
+------+--------------------------------------------------------+
|apple |{"name":"apple","color":"red","taste":"sweet"}          |
|banana|{"name":"banana","color":"yellow","taste":"sweet"}      |
|cherry|{"name":"cherry","color":"red","taste":"sweet and tart"}|
+------+--------------------------------------------------------+

Overview

This package provides a vectorized interface for the OpenAI API, enabling you to process multiple inputs with a single API call instead of sending requests one by one. This approach helps reduce latency and simplifies your code.

Additionally, it integrates effortlessly with Pandas DataFrames and Apache Spark UDFs, making it easy to incorporate into your data processing pipelines.

Features

• Vectorized API requests for processing multiple inputs at once.
• Seamless integration with Pandas DataFrames.
• A UDF builder for Apache Spark.
• Compatibility with multiple OpenAI clients, including Azure OpenAI.

Requirements

• Python 3.10 or higher

Installation

Install the package with:

pip install openaivec

If you want to uninstall the package, you can do so with:

pip uninstall openaivec

Basic Usage

import os
from openai import OpenAI
from openaivec import VectorizedOpenAI


# Initialize the vectorized client with your system message and parameters
client = VectorizedOpenAI(
    client=OpenAI(),
    temperature=0.0,
    top_p=1.0,
    model_name="<your-model-name>",
    system_message="Please answer only with 'xx family' and do not output anything else."
)

result = client.predict(["panda", "rabbit", "koala"])
print(result)  # Expected output: ['bear family', 'rabbit family', 'koala family']

See examples/basic_usage.ipynb for a complete example.

Using with Pandas DataFrame

openaivec.pandas_ext extends pandas.Series functions with accessor ai.predict or ai.embed.

import pandas as pd
from openaivec import pandas_ext

df = pd.DataFrame({"name": ["panda", "rabbit", "koala"]})

df.assign(
    kind=lambda df: df.name.ai.predict("gpt-4o", "Answer only with 'xx family' and do not output anything else.")
)

Example output:

name	kind
panda	bear family
rabbit	rabbit family
koala	koala family

Using with Apache Spark UDF

Below is an example showing how to create UDFs for Apache Spark using the provided UDFBuilder. This configuration is intended for use with Azure OpenAI or OpenAI.

from openaivec.spark import UDFBuilder

udf = UDFBuilder.of_azureopenai(
    api_key="<your-api-key>",
    api_version="2024-10-21",
    endpoint="https://<your_resource_name>.openai.azure.com",
    model_name="<your_deployment_name>"
)

# Register UDFs (e.g., to extract flavor or product type from product names)
spark.udf.register("parse_taste", udf.completion("""
- Extract flavor-related information from the product name. Return only the concise flavor name with no extra text.
- Minimize unnecessary adjectives related to the flavor.
    - Example:
        - Hokkaido Milk → Milk
        - Uji Matcha → Matcha
"""))

# Register UDFs (e.g., to extract product type from product names)
spark.udf.register("parse_product", udf.completion("""
- Extract the type of food from the product name. Return only the food category with no extra text.
- Example output:
    - Smoothie
    - Milk Tea
    - Protein Bar
"""))

You can then use the UDFs in your Spark SQL queries as follows:

SELECT id,
       product_name,
       parse_taste(product_name)   AS taste,
       parse_product(product_name) AS product
FROM product_names;

Example Output:

id	product_name	taste	product
4414732714624	Cafe Mocha Smoothie (Trial Size)	Mocha	Smoothie
4200162318339	Dark Chocolate Tea (New Product)	Chocolate	Tea
4920122084098	Cafe Mocha Protein Bar (Trial Size)	Mocha	Protein Bar
4468864478874	Dark Chocolate Smoothie (On Sale)	Chocolate	Smoothie
4036242144725	Uji Matcha Tea (New Product)	Matcha	Tea
4847798245741	Hokkaido Milk Tea (Trial Size)	Milk	Milk Tea
4449574211957	Dark Chocolate Smoothie (Trial Size)	Chocolate	Smoothie
4127044426148	Fruit Mix Tea (Trial Size)	Fruit	Tea
...	...	...	...

Building Prompts

Building prompt is a crucial step in using LLMs. In particular, providing a few examples in a prompt can significantly improve an LLM’s performance, a technique known as "few-shot learning." Typically, a few-shot prompt consists of a purpose, cautions, and examples.

FewShotPromptBuilder is a class that helps you build a few-shot learning prompt with simple interface.

Basic Usage

FewShotPromptBuilder requires simply a purpose, cautions, and examples, and build method will return rendered prompt with XML format.

Here is an example:

from openaivec.prompt import FewShotPromptBuilder

prompt: str = (
    FewShotPromptBuilder()
    .purpose("Return the smallest category that includes the given word")
    .caution("Never use proper nouns as categories")
    .example("Apple", "Fruit")
    .example("Car", "Vehicle")
    .example("Tokyo", "City")
    .example("Keiichi Sogabe", "Musician")
    .example("America", "Country")
    .build()
)
print(prompt)

The output will be:

<Prompt>
    <Purpose>Return the smallest category that includes the given word</Purpose>
    <Cautions>
        <Caution>Never use proper nouns as categories</Caution>
    </Cautions>
    <Examples>
        <Example>
            <Source>Apple</Source>
            <Result>Fruit</Result>
        </Example>
        <Example>
            <Source>Car</Source>
            <Result>Vehicle</Result>
        </Example>
        <Example>
            <Source>Tokyo</Source>
            <Result>City</Result>
        </Example>
        <Example>
            <Source>Keiichi Sogabe</Source>
            <Result>Musician</Result>
        </Example>
        <Example>
            <Source>America</Source>
            <Result>Country</Result>
        </Example>
    </Examples>
</Prompt>

Improve with OpenAI

For most users, it can be challenging to write a prompt entirely free of contradictions, ambiguities, or redundancies. FewShotPromptBuilder provides an improve method to refine your prompt using OpenAI's API.

improve method will try to eliminate contradictions, ambiguities, and redundancies in the prompt with OpenAI's API, and iterate the process up to max_iter times.

Here is an example:

from openai import OpenAI
from openaivec.prompt import FewShotPromptBuilder

client = OpenAI(...)
model_name = "<your-model-name>"
improved_prompt: str = (
    FewShotPromptBuilder()
    .purpose("Return the smallest category that includes the given word")
    .caution("Never use proper nouns as categories")
    # Examples which has contradictions, ambiguities, or redundancies
    .example("Apple", "Fruit")
    .example("Apple", "Technology")
    .example("Apple", "Company")
    .example("Apple", "Color")
    .example("Apple", "Animal")
    # improve the prompt with OpenAI's API, max_iter is number of iterations to improve the prompt.
    .improve(client, model_name, max_iter=5)
    .build()
)
print(improved_prompt)

Then we will get the improved prompt with extra examples, improved purpose, and cautions:

<Prompt>
    <Purpose>Classify a given word into its most relevant category by considering its context and potential meanings.
        The input is a word accompanied by context, and the output is the appropriate category based on that context.
        This is useful for disambiguating words with multiple meanings, ensuring accurate understanding and
        categorization.
    </Purpose>
    <Cautions>
        <Caution>Ensure the context of the word is clear to avoid incorrect categorization.</Caution>
        <Caution>Be aware of words with multiple meanings and provide the most relevant category.</Caution>
        <Caution>Consider the possibility of new or uncommon contexts that may not fit traditional categories.</Caution>
    </Cautions>
    <Examples>
        <Example>
            <Source>Apple (as a fruit)</Source>
            <Result>Fruit</Result>
        </Example>
        <Example>
            <Source>Apple (as a tech company)</Source>
            <Result>Technology</Result>
        </Example>
        <Example>
            <Source>Java (as a programming language)</Source>
            <Result>Technology</Result>
        </Example>
        <Example>
            <Source>Java (as an island)</Source>
            <Result>Geography</Result>
        </Example>
        <Example>
            <Source>Mercury (as a planet)</Source>
            <Result>Astronomy</Result>
        </Example>
        <Example>
            <Source>Mercury (as an element)</Source>
            <Result>Chemistry</Result>
        </Example>
        <Example>
            <Source>Bark (as a sound made by a dog)</Source>
            <Result>Animal Behavior</Result>
        </Example>
        <Example>
            <Source>Bark (as the outer covering of a tree)</Source>
            <Result>Botany</Result>
        </Example>
        <Example>
            <Source>Bass (as a type of fish)</Source>
            <Result>Aquatic Life</Result>
        </Example>
        <Example>
            <Source>Bass (as a low-frequency sound)</Source>
            <Result>Music</Result>
        </Example>
    </Examples>
</Prompt>

Using with Microsoft Fabric

Microsoft Fabric is a unified, cloud-based analytics platform that seamlessly integrates data engineering, warehousing, and business intelligence to simplify the journey from raw data to actionable insights.

This section provides instructions on how to integrate and use vectorize-openai within Microsoft Fabric. Follow these steps:

1. Create an Environment in Microsoft Fabric:

   • In Microsoft Fabric, click on New item in your workspace.
   • Select Environment to create a new environment for Apache Spark.
   • Determine the environment name, eg. openai-environment.
   • Figure: Creating a new Environment in Microsoft Fabric.

2. Add openaivec to the Environment from Public Library

   • Once your environment is set up, go to the Custom Library section within that environment.
   • Click on Add from PyPI and search for latest version of openaivec.
   • Save and publish to reflect the changes.
   • Figure: Add openaivec from PyPI to Public Library

3. Use the Environment from a Notebook:

   • Open a notebook within Microsoft Fabric.

   • Select the environment you created in the previous steps.

   • Figure: Using custom environment from a notebook.

   • In the notebook, import and use openaivec.spark.UDFBuilder as you normally would. For example:

     from openaivec.spark import UDFBuilder
     
     udf = UDFBuilder(
         api_key="<your-api-key>",
         api_version="2024-10-21",
         endpoint="https://<your-resource-name>.openai.azure.com",
         model_name="<your-deployment-name>"
     )

Following these steps allows you to successfully integrate and use vectorize-openai within Microsoft Fabric.

Contributing

We welcome contributions to this project! If you would like to contribute, please follow these guidelines:

1. Fork the repository and create your branch from main.
2. If you've added code that should be tested, add tests.
3. Ensure the test suite passes.
4. Make sure your code lints.

Installing Dependencies

To install the necessary dependencies for development, run:

poetry install --dev

Code Formatting

To reformat the code, use the following command:

poetry run ruff check . --fix

Community

Join our Discord community for developers: https://discord.gg/vbb83Pgn