Skip to content

breakbotz/artkit

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

17 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

ARTKIT logo

Automated Red Teaming (ART) and testing toolkit

ARTKIT is a Python framework developed by BCG X for automating prompt-based testing and evaluation of Gen AI applications.

Getting started

Introduction

ARTKIT is a Python framework for developing automated end-to-end testing and evaluation pipelines for Gen AI applications. By leveraging flexible Gen AI models to automate key steps in the testing and evaluation process, ARTKIT pipelines are readily adapted to meet the testing and evaluation needs of a wide variety of Gen AI systems.

ARTKIT pipeline schematic

ARTKIT also supports automated multi-turn conversations between a challenger bot and a target system. Issues and vulnerabilities are more likely to arise after extended interactions with Gen AI systems, so multi-turn testing is critical for interactive applications.

We recommend starting with our User Guide to learn the core concepts and functionality of ARTKIT. Visit our Examples to see how ARTKIT can be used to test and evaluate Gen AI systems for:

  1. Q&A Accuracy:
  2. Upholding Brand Values:
    • Implement persona-based testing to simulate diverse users interacting with your system and evaluate system responses for brand conformity.
  3. Equitability:
    • Run a counterfactual experiment by systematically modifying demographic indicators across a set of documents and statistically evaluate system responses for undesired demographic bias.
  4. Safety:
  5. Security:

These are just a few examples of the many ways ARTKIT can be used to test and evaluate Gen AI systems for proficiency, equitability, safety, and security.

Key Features

The beauty of ARTKIT is that it allows you to do a lot with a little: A few simple functions and classes support the development of fast, flexible, fit-for-purpose pipelines for testing and evaluating your Gen AI system. Key features include:

  • Simple API: ARTKIT provides a small set of simple but powerful functions that support customized pipelines to test and evaluate virtually any Gen AI system.
  • Asynchronous: Leverage asynchronous processing to speed up processes that depend heavily on API calls.
  • Caching: Manage development costs by caching API responses to reduce the number of calls to external services.
  • Model Agnostic: ARTKIT supports connecting to major Gen AI model providers and allows users to develop custom model classes to connect to any GenAI service.
  • End-to-End Pipelines: Build end-to-end flows to generate test prompts, interact with a target system (i.e., system being tested), perform quantitative evaluations, and structure results for reporting.
  • Multi-Turn Conversations: Create automated interactive dialogs between a target system and an LLM persona programmed to interact with the target system in pursuit of a specific goal.
  • Robust Data Flows: Automatically track the flow of data through testing and evaluation pipelines, facilitating full traceability of data lineage in the results.
  • Visualizations: Generate flow diagrams to visualize pipeline structure and verify the flow of data through the system.

Note

ARTKIT is designed to be customized by data scientists and engineers to enhance human-in-the-loop testing and evaluation. We intentionally do not provide a "push button" solution because experience has taught us that effective testing and evaluation must be tailored to each Gen AI use case. Automation is a strategy for scaling and accelerating testing and evaluation, not a substitute for case-specific risk landscape mapping, domain expertise, and critical thinking.

Supported Model Providers

ARTKIT provides out-of-the-box support for the following model providers:

To connect to other services, users can develop custom model classes.

Installation

ARTKIT supports both PyPI and Conda installations. We recommend installing ARTKIT in a dedicated virtual environment.

Pip

MacOS and Linux:

python -m venv artkit
source artkit/bin/activate
pip install artkit

Windows:

python -m venv artkit
artkit\Scripts\activate.bat
pip install artkit

Conda

conda install -c conda-forge artkit

Optional dependencies

To enable visualizations of pipeline flow diagrams, install GraphViz and ensure it is in your system's PATH variable:

Environment variables

Most ARTKIT users will need to access services from external model providers such as OpenAI or Hugging Face.

Our recommended approach is:

  1. Install python-dotenv using pip:
pip install python-dotenv

or conda:

conda install -c conda-forge python-dotenv
  1. Create a file named .env in your project root.
  2. Add .env to your .gitignore to ensure it is not committed to your Git repo.
  3. Define environment variables inside .env, for example, API_KEY=your_api_key
  4. In your Python scripts or notebooks, load the environmental variables with:
from dotenv import load_dotenv
load_dotenv()

# Verify that the environment variable is loaded
import os
os.getenv('YOUR_API_KEY')

The ARTKIT repository includes an example file called .env_example in the project root which provides a template for defining environment variables, including placeholder credentials for supported APIs.

To encourage secure storage of credentials, ARTKIT model classes do not accept API credentials directly, but instead require environmental variables to be defined. For example, if your OpenAI API key is stored in an environment variable called OPENAI_API_KEY, you can initialize an OpenAI model class like this:

import artkit.api as ak

ak.OpenAIChat(
    model_id="gpt-4o",
    api_key_env="OPENAI_API_KEY"
    )

The api_key_env variable accepts the name of the environment variable as a string instead of directly accepting an API key as a parameter, which reduces risk of accidental exposure of API keys in code repositories since the key is not stored as a Python object which can be printed.

Quick Start

The core ARTKIT functions are:

  1. run: Execute one or more pipeline steps
  2. step: A single pipeline step which produces a dictionary or an iterable of dictionaries
  3. chain: A set of steps that run in sequence
  4. parallel: A set of steps that run in parallel

Below, we develop a simple example pipeline with the following steps:

  1. Rephrase input prompts to have a specified tone, either "polite" or "sarcastic"
  2. Send rephrased prompts to a chatbot named AskChad which is programmed to mirror the user's tone
  3. Evaluate the responses according to a "sarcasm" metric

To begin, import artkit.api and set up a session with the OpenAI GPT-4o model. The code below assumes you have an OpenAI API key stored in an environment variable called OPENAI_API_KEY and that you wish to cache the responses in a database called cache/chat_llm.db.

import artkit.api as ak

# Set up a chat system with the OpenAI GPT-4o model
chat_llm = ak.CachedChatModel(
    model=ak.OpenAIChat(model_id="gpt-4o"),
    database="cache/chat_llm.db"
)

Next, define a few functions that will be used as pipeline steps. ARTKIT is designed to work with asynchronous generators to allow for asynchronous processing, so the functions below are defined with async, await, and yield keywords.

# A function that rephrases input prompts to have a specified tone
async def rephrase_tone(prompt: str, tone: str, llm: ak.ChatModel):

    response = await llm.get_response(
        message = (
            f"Your job is to rephrase in input question to have a {tone} tone.\n"
            f"This is the question you must rephrase:\n{prompt}"
        )
    )

    yield {"prompt": response[0], "tone": tone}


# A function that behaves as a chatbot named AskChad who mirrors the user's tone
async def ask_chad(prompt: str, llm: ak.ChatModel):

    response = await llm.get_response(
        message = (
            "You are AskChad, a chatbot that mirrors the user's tone. "
            "For example, if the user is rude, you are rude. "
            "Your responses contain no more than 10 words.\n"
            f"Respond to this user input:\n{prompt}"
        )
    )

    yield {"response": response[0]}


# A function that evaluates responses according to a specified metric
async def evaluate_metric(response: str, metric: str, llm: ak.ChatModel):

    score = await llm.get_response(
        message = (
            f"Your job is to evaluate prompts according to whether they are {metric}. "
            f"If the input prompt is {metric}, return 1, otherwise return 0.\n"
            f"Please evaluate the following prompt:\n{response}"
        )
    )

    yield {"evaluation_metric": metric, "score": int(score[0])}

Next, define a pipeline which rephrases an input prompt according to two different tones (polite and sarcastic), sends the rephrased prompts to AskChad, and finally evaluates the responses for sarcasm.

pipeline = (
    ak.chain(
        ak.parallel(
            ak.step("tone_rephraser", rephrase_tone, tone="POLITE", llm=chat_llm),
            ak.step("tone_rephraser", rephrase_tone, tone="SARCASTIC", llm=chat_llm),
        ),
        ak.step("ask_chad", ask_chad, llm=chat_llm),
        ak.step("evaluation", evaluate_metric, metric="SARCASTIC", llm=chat_llm)
    )
)

pipeline.draw()

sphinx/source/_images/quick_start_flow_diagram.png

Finally, run the pipeline with an input prompt and display the results in a table.

# Input to run through the pipeline
prompt = {"prompt": "What is a fun activity to do in Boston?"}

# Run pipeline
result = ak.run(steps=pipeline, input=prompt)

# Convert results dictionary into a multi-column dataframe
result.to_frame()

sphinx/source/_images/quick_start_results.png

From left to right, the results table shows:

  1. input: The original input prompt
  2. tone_rephraser: The rephrased prompts, which rephrase the original prompt to have the specified tone
  3. ask_chad: The response from AskChad, which mirrors the tone of the user
  4. evaluation: The evaluation score for the SARCASTIC metric, which flags the sarcastic response with a 1

For a complete introduction to ARTKIT, please visit our User Guide and Examples.

Contributing

Contributions to ARTKIT are welcome and appreciated! Please see the Contributor Guide section for information.

License

This project is licensed under Apache 2.0, allowing free use, modification, and distribution with added protections against patent litigation. See the LICENSE file for more details or visit Apache 2.0.

BCG X

BCG X is the tech build and design unit of Boston Consulting Group.

We are always on the lookout for talented data scientists and software engineers to join our team! Visit BCG X Careers to learn more.

About

Automated prompt-based testing and evaluation of Gen AI applications

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 98.1%
  • Python 1.9%