awesome-deepseek-integration/docs/promptfoo

promptfoo

promptfoo is an open-source framework for testing and evaluating LLM outputs. It helps you compare DeepSeek models with other LLMs (like o1, GPT-4o, Claude 3.5, Llama3.3, and Gemini) and test LLMs and LLM applications for security vulnerabilities. You can:

• Run side-by-side comparisons between models
• Check output quality and consistency
• Generate test reports

Setup

1. Install promptfoo:

npm install -g promptfoo
# or
brew install promptfoo

2. Configure API keys:

export DEEPSEEK_API_KEY=your_api_key
# Add other API keys as needed

Quick Start

Create a configuration file promptfooconfig.yaml:

providers:
  - deepseek:deepseek-reasoner # DeepSeek-R1
  - openai:o1

prompts:
  - 'Solve this step by step: {{math_problem}}'

tests:
  - vars:
      math_problem: 'What is the derivative of x^3 + 2x with respect to x?'
    assert:
      - type: contains
        value: '3x^2' # Check for correct answer
      - type: llm-rubric
        value: 'Response shows clear steps'
      - type: cost
        threshold: 0.05 # Maximum cost per request

Run tests:

promptfoo eval

View results in your browser:

promptfoo view

Example Project

Check out our example that compares r1 and o1 on MMLU.

Resources

• Documentation
• GitHub Repository
• Community Discord