/contextcheck

MIT-licensed Framework for LLMs, RAGs, Chatbots testing. Configurable via YAML and integrable into CI pipelines for automated testing.

Primary LanguagePythonMIT LicenseMIT

ContextCheck Logo

ContextCheck: Test your LLMs, RAGs and Chatbots

ContextCheck is a MIT-licensed framework for testing, evaluating LLMs, RAGs, and chatbots.

ContextCheck is an open-source framework designed to evaluate, test, and validate large language models (LLMs), Retrieval-Augmented Generation (RAG) systems, and AI chatbots. It offers AI testing tools to automatically generate queries, request completions, detect regressions, perform penetration tests, and assess hallucinations, ensuring the robustness and reliability of these systems. ContextCheck is fully configurable via YAML and integrates seamlessly into continuous integration (CI) pipelines for efficient prompt testing and automated validation.

Table of Contents

Features

  • Simple test scenario definition using human-readable .yaml files
  • Flexible endpoint configuration for OpenAI, HTTP, and more
  • Customizable JSON request/response models
  • Support for variables and Jinja2 templating in YAML files
  • Response validation options, including heuristics, LLM-based judgment, and human labeling
  • Enhanced output formatting with the rich package for clear, readable displays

LLM Prompts Testing

An LLM prompt is a specific input or query provided to a language model to generate a response based on the model's training data. ContextCheck provides an easy-to-use tool for fine-tuning and testing multiple prompts at once.

LLM Prompts Testing Process

The prompts testing workflow consists of the following steps:

  1. Choose an LLM: Select the provider and model version you want to test your prompts with.
  2. Write test scenario(s).

Each test scenario requires you to:

  1. Define LLM Prompts: Create the prompts you want to test.
  2. Write Tests: Create test assertions that verify the effectiveness of your prompts. These tests check if the model's responses meet your expectations.
  3. Execute Tests: Run your tests to see which prompts pass or fail.
  4. Fine-Tune Prompts: If any tests fail, adjust your prompts and re-run the tests. Repeat this process until all tests pass or you’re satisfied with the results.

ContextCheck allows you to execute individual or batch test scenarios.

You can test your prompts using either deterministic (rule-based) metrics, LLM-based metrics, or a combination of both, depending on your needs. To better understand the available metrics and how to use them, refer to the metrics documentation.

For technical details how to configure and write tests, refer to How to configure test scenario.

Testing Prompts Across Different Models

Based on your particular requirements, you might want check how your prompts will perform against various LLM models. Once scenario, or multiple scenarios, are ready you can switch LLM model and run tests against it.

The LLM landscape changes rapidly. When new models or model versions become available, and you want to test how they perform against your prompts, simply change test scenario configuration and execute you tests.

Note that in the case of prompt validation, not all LLM-based metrics are applicable. For example, the hallucination metric requires reference documents to verify if the answer is based solely on the reference or if it is fabricated.

RAG (Retrieval Augmented Generation) System Testing

This scenario is similar to the "prompts validation" but instead of testing LLMs, the goal is to validate the performance of the RAG system API endpoint. ContextCheck sends a query to the API endpoint, receives a response from the RAG system, and validates that answer.

As the name suggests, RAG systems generate their answers based on the content retrieved from external sources or databases, combining retrieval and generation to provide more accurate and contextually relevant responses.

RAG systems can range from simple implementations that query a single document collection to complex architectures that integrate multiple knowledge sources, language models, and reasoning components, differing in their retrieval mechanisms, the sophistication of their knowledge integration, and their ability to perform multi-step reasoning or handle diverse types of information.

Depending on your needs, ContextCheck can test only the response from the RAG system or also examine its internal workings. For instance, using the hallucination metric, you can retrieve the reference documents the RAG system used to generate the response and ensure that no additional (hallucinated) information was introduced by the LLM.

Currently, ContextCheck only supports fetching reference documents. However, future updates will enable users to extract and test inputs or outputs from any component of a RAG system. These features will be implemented with minimal changes required to the RAG system's codebase.

RAG System Testing Process

The prompts testing workflow consists of the following steps:

  1. Configure: Both ContextCheck and (depending on particular use-case) RAG system have to be configured. Refer to details in chapter RAG Configuration.
  2. Write test scenario(s).

Each test scenario requires you to:

  1. Define queries: Write queries that will be sent to your RAG system.
  2. Write Tests: Create tests that verify the effectiveness of your queries. Additionally, you can use for your tests the RAG metadata, like for example documents that were pulled from RAG's database and used to generate the response. Similarly to LLM prompts testing process, these tests check if the model's responses meet your expectations.
  3. Execute Tests: Run your tests to see which queries pass or fail.
  4. Fine-Tune tests: If any tests fail, adjust them accordingly and re-run. Repeat this process until all tests pass or you’re satisfied with the results.

Test-driven RAG system development

The process described above enables a rigorous approach to making changes in the system. After introducing a change, you can run the entire test suite to check if it caused any regressions in other parts of the system. When integrated with a CI/CD workflow, this process can protect the production environment from bugs or regressions. It ensures that changes are thoroughly tested before being deployed to production.

Metrics

To write test assertions, you'll use specific, preconfigured metrics. ContextCheck implements both deterministic (rule-based) metrics and LLM-based metrics. When creating your tests, you can choose to use deterministic metrics, LLM-based metrics, or a combination of both, depending on your specific needs and the nature of your test scenarios. To better understand the available metrics and how to use them, refer to the metrics documentation.

For technical details how to configure and write tests, refer to How to configure test scenario.

Installation

For Users

Install the package directly from PyPI using pip:

pip install ccheck

After installation, you can access the ccheck CLI command:

ccheck --help

This will display all available options and help you get started with using ContextCheck.

For Developers

If you wish to contribute to the project or modify it for your own use, you can set up a development environment using Poetry.

  1. Fork your own copy of Addepto/contextcheck on GitHub.
  2. Clone the Repository:
git clone https://github.com/<your_username>/contextcheck.git
cd contextcheck
  1. Ensure you have Poetry installed.
  2. Install Dependencies:
poetry install
  1. Activate the Virtual Environment:
poetry shell
  1. Activate the ccheck CLI command using:
poetry run ccheck --help

Tutorial

Please refer to examples/ folder for the tutorial.

CLI Features

Output Test Results to Console

  • Run a single scenario and output results to the console: 
    ccheck --output-type console --filename path/to/file.yaml
  • Run multiple scenarios and output results to the console: 
    ccheck --output-type console --filename path/to/file.yaml path/to/another_file.yaml

Running in CI/CD

To automatically stop the CI/CD process if any tests fail, add the --exit-on-failure flag. Failed test will cause the script to exit with code 1:

ccheck --exit-on-failure --output-type console --folder my_tests

Use env variable OPENAI_API_KEY to be able to run:

  • tests/scenario_openai.yaml
  • tests/scenario_defaults.yaml

Contributing

Contributions are welcomed!

Running Tests

To run tests:

poetry run pytest tests/

To include tests which require calling LLM APIs (currently OpenAI and Ollama), run one of:

poetry run pytest --openai          # includes tests that use OpenAI API
poetry run pytest --ollama          # includes tests that use Ollama API
poetry run pytest --openai --ollama # includes tests that use both OpenAI and Ollama API

Acknowledgments

Made with ❤️ by the Addepto Team

ContextCheck is an extension of the ContextClue product, created by the Addepto team. This project is the result of our team’s dedication, combining innovation and expertise.

Addepto Team:

  • Radoslaw Bodus
  • Bartlomiej Grasza
  • Volodymyr Kepsha
  • Vadym Mariiechko
  • Michal Tarkowski

Like what we’re building? ⭐ Give it a star to support its development!

License

This project is licensed under the MIT License - see the LICENSE file for details