Skip to content

Latest commit

 

History

History
253 lines (173 loc) · 14.5 KB

README.md

File metadata and controls

253 lines (173 loc) · 14.5 KB

Prompt Wizard

Open In Colab

Description

Prompt Wizard is a package for evaluating custom prompts using various evaluation methods. It allows you to provide your own prompts or generate them automatically and then obtain the results in a JSON file.

Features

  • Evaluate custom prompts using different evaluation methods.
  • Automatically generate prompts for evaluation.
  • Iterates over your best performing prompts to get better prompts.
  • Save the results obtained in a JSON file.
  • Use the library for a project in which the tools that promptwizard provides are useful.

Installation

To use Prompt Wizard you need to install the package and all its dependencies using pip

pip install promptwizard

Alternatively, clone the repository using git clone https://github.com/Leniolabs/promptwiz.git

Setup

To run Prompt Wizard, you will need to set up and specify your OpenAI API key. You can generate one at https://platform.openai.com/account/api-keys. After you obtain an API key, specify it using the OPENAI_API_KEY environment variable. Please be aware of the costs associated with using the API when running evals.

Authentication

Before using Prompt Wizard you will need to have your environment variables defined. You have two valid options, have your OPENAI_API_KEY defined in an .env in the correct folder, or, if you decide to use azure you need to define your OPENAI_API_TYPE as azure and your OPENAI_API_BASE and OPENAI_API_VERSION correctly in your .env in addition to the OPENAI_API_KEY.

Usage

Prompt Evaluation

You have two usage alternatives.

If you want to use YAML files:

  1. Make sure you have the YAML file with the prompts you want to evaluate. The YAML file should follow the proper structure.

  • Run the package with the YAML file as an argument if you have your .env file in the same directory as the package:
promptwizard YAML_FILE_PATH
  • Run the package with the YAML file as an argument with the path of your .env file as shown below
promptwizard YAML_FILE_PATH --env_path .env_FILE_PATH

Respond 'Y' when asked if you want to continue.

  1. The evaluation result will be saved in an output.json file in the same folder as the YAML file. If you choose the Elo method for prompt evaluation, a scatter plot scatter_plot.png will also be saved in the same folder as the YAML file. A larger number of files will also be generated if you have indicated in your yaml file that you want to perform iterations.

Automatic Prompt Generation

  1. If the "prompts" variable is not defined in the YAML file, the program will automatically generate prompts for evaluation.

  2. Run the package passing your YAML file as parameter:

promptwizard YAML_FILE_PATH

Respond 'Y' when asked if you want to continue.

  1. The evaluation result will be saved in an output.json file in the same folder as the YAML folder. If you choose the Elo method for prompt evaluation, a scatter plot scatter_plot.png will also be saved in the same folder as the YAML file.

You can also use it in your Python script:

import promptwizard

And use the various functions that PromptWizard can provide you with, such as for example:

# Example of using PromptWizard
from promptwizard import prompt_generation

test_cases = [
    {'input': 'How do you make a classic spaghetti carbonara?', 'output': 'REPLY'},
    {'input': "What is John Smith's phone number?", 'output': 'NOT_REPLY'},
]
description = "Decide whether the question should be answered or not." #A short description of the type of task for the test cases.
system_gen_prompt = """Your job is to generate system prompts for GPT, given a description of the use-case and some test cases.
In your generated prompt, you should describe how the AI should behave in plain English. Include what it will see, and what it's allowed to output. Be creative with prompts to get the best possible results. The AI knows it's an AI -- you don't need to tell it this.
Remember that the prompt should only allow the AI to answer the answer and nothing else. No explanation is necessary.
You will be graded based on the performance of your prompt... but don't cheat! You cannot include specifics about the test cases in your prompt. Any prompts with examples will be disqualified. I repeat, do not include the test cases.
Most importantly, output NOTHING but the prompt. Do not include anything else in your message.""" #Here you have to indicate to the LLM how your generated prompts should be. This example is useful if you later want to use the equals evaluation method.
# Create 4 prompts.
prompts = prompt_generation.generate_candidate_prompts(system_gen_prompt, test_cases, description)[0]

Prompt Iteration

If you want, you can also specify the number of iterations you want to perform on your provided prompts or the ones that will be generated automatically to obtain prompts that achieve optimal behavior for the language model. On the other hand, the use of functions to iterate prompts for your python script is available:

from promptwizard.prompt_generation import iteration

results = iteration.iterations(test_cases, method='Elo', prompts=old_prompts, number_of_prompts=3)

YAML Files

We provide you an explanation of the valid structure of your YAML files and certain limitations for some variables within it. We recommend that you read it carefully before running an evaluation.

The following is the structure that your YAML files should have.

test:

    cases: """ Here, you have to put the test cases you are going to use to evaluate your prompts. If you are going to use the
        Elo method to evaluate them, it should be just a list of strings. If you are going to use the methods classification, 
        equal or includes, it should be a list of tuples with two elements, where the first element is the test case and the 
        second element is the correct response to the test. Remember that if you decide to use classification, only a boolean
        value is allowed as a response. the form of your test cases has to be, in case of selecting the Elo method:
            -'Test1'
            -'Test2'...
        If you choose the methods Classification, Equals, Includes, Semantic Similarity or LogProbs they must be of the form:
            -input: 'Test1'
            output: 'Answer1'
            -input: 'Test2'
            output: 'Answer2'
        In case the method is Function Calling:
            -input: 'Test1'
            output1: 'name_function'
            output2: 'variable'
            -input: 'Test2'
            output1: 'name_function'
            output2: 'variable'
        If you choose Code Generation:
            - input: 'Test1'
            arguments: (arg1,) in case there is only one argument, (arg1, arg2,...) in case there are more than one argument.
            output: res
        and finally if you choose JSON Validation:
            - input: 'Test1'
            output: json_output"""

    description: """Here is the description of the type of task that summarizes the test cases. You only have to use this field if 
        you are going to use the 'Elo' method"""
    method: """Here, you select the evaluation method for your prompts. You must choose between 'Elo',
        'Classification', 'Equals', 'Includes', 'Function Calling', 'Code Generation' 'JSON Validation', 'Semantic Similarity' and 'LogProbs'."""

    model:
        name: """The name of the GPT model you will use to evaluate the prompts."""
        temperature: """The temperature of the GPT model you will use to evaluate the prompts."""
        max_tokens: """The maximum number of tokens you will allow the GPT model to use to generate the response to the test."""

    functions: """This field must only be filled out in case the 'Function Calling' method is intended to be used.
    If another method is used, it must not be filled out. The structure is a JSON object. Let's break down the different components:

            - Function Name (name): This is the identifier used to refer to this function within the context of your code.

            - Function Description (description): A brief description of what the function does.

            - Function Parameters (parameters): This section defines the input parameters that the function accepts.

                - Type (type): The type of the parameter being defined.

                - Properties (properties): This is an object containing properties that the input parameter object should have.

                    - File Type (file_type): This is a property of the parameter object.

                    - Enum (enum): An enumeration of allowed values for the 'file_type' property. (optional)

                    - Description (description): A description of what the 'file_type' property represents.

                - Required (required): An array listing the properties that are required within the parameter object. (optional)"""

    function_call: """This field must only be filled out in case the 'Function Calling' method is intended to be 
            used. If another method is used, it must not be filled out."""

prompts: """You have two options, either provide your list of prompts or generate them following the instructions below."""

    list: """A list of prompts you want to evaluate. If you want to generate them with the prompt generator, don't use this field.
        Please provide a minimum number of 4 prompts. Your prompts must be listed as follows:
            - 'Prompt1'
            - 'Prompt2'..."""

    generation:

        number: """The number of prompts you are going to evaluate. You need to provide this key value only if you are going to generate the prompts. Indicate the quantity of prompts you want to generate. Please provide a minimum number of 4 prompts. If you do not define this key by default, 4 prompts will be created."""

        constraints: """If you are going to generate prompts, this optional feature allows you to add special characteristics to the prompts that will be generated. For example, if you want prompts with a maximum length of 50 characters, simply complete with 'Generate prompts with a maximum length of 50 characters'. If you don't want to use it, you don't need to have this key defined."""

        description: """Here is the description of the type of task that summarizes the test cases. If you use the 'Elo' method you mustn't use this field."""

        best_prompts: """The number of prompts you want to iterate over and on which you want to highlight the final results. the value must be between 2 and the number of prompts you provide (or generate) minus one. If you do not define this value the default value will be 2."""

        model:

            name: """The name of the GPT model you will use to generate the prompts."""
            temperature: """The temperature of the GPT model you will use to generate the prompts."""
            max_tokens: """The maximum number of tokens you will allow the GPT model to use to generate your prompts."""

    iterations:
        number: """The number of iterations you want to perform on the best prompts obtained in your initial testing to arrive at 
            prompts with better final results. If you don't want to try alternatives combining your best prompts just put 0."""

        best_percentage: """Number between 0 and 100 indicating that iterations should be stopped if all 'best_prompts' equaled or exceeded the indicated accuracy. If this value is not defined, it will default to 100."""

        model:

            name: """The name of the GPT model you will use to generate the prompts."""
            temperature: """The temperature of the GPT model you will use to generate the prompts."""
            max_tokens: """The maximum number of tokens you will allow the GPT model to use to generate your prompts."""
            You can not define these variables for 'model' in case you want to keep the same variables that were used in 'generation', in case the 'generation' field has not been used it will take the following default values:
            name: 'gpt-4
            temperature: 0.6
            max_tokens: 300"""
timeout: """Timeout set for an API request. This time limit indicates how long the client should wait to receive a response before the request expires."""

n_retries: """Number of attempts that will be automatically made to resend an API request in case the initial request fails."""

In case the YAML file you wish to evaluate has errors in its structure, don't worry. Prior to being assessed by the prompt engineer, your file will be validated, and you will receive a notification indicating where you need to make corrections to it in order to be evaluated successfully.

Special features

Remember that when you generate your prompts you can use the constraints key to explicitly request that the prompts you are going to generate have a special characteristic, for example, 'generate prompts of a length not exceeding 20 words'.

Cost and tokens

If you want to know how much it will cost to run your evaluation, simply enter:

promptwizard YAML_FILE_PATH

and simply answer 'N' when asked if you want to continue.

Otherwise, respond 'Y' and run your evaluation and you'll receive the approximate cost, along with the real final cost at the end. In the final JSON file, in addition to seeing the top prompts with the best results, you will also have this same information about the costs and the number of tokens effectively consumed for both GPT-3.5-turbo and GPT-4.

ALternatively, you can do the following in your Python script:

from promptwizard.approximate_cost import cost

print(cost.approximate_cost(test_cases, method, prompts_value))

and you're going to see an approximate cost for your possible evaluation.

Usage examples

If you want to see usage examples, we provide the following Colab notebook for you to explore the different ways you can use PromptWizard. (https://colab.research.google.com/drive/1iW2y43923vECOhKPuhoGENwY1y81Rw8i?usp=sharing)

Creators

promptwizard is crafted with love by Leniolabs and a growing community of contributors. We build digital experiences with your ideas. Get in touch! Also, if you have any questions or feedback about promptwizard, please feel free to contact us at [email protected]. We'd love to hear from you!