Promptfoo — LLM Eval & Red-Team Testing Framework

What it is

Promptfoo is an open-source CLI and library for evaluating LLM outputs systematically. It runs your prompts against multiple models, scores the outputs using custom assertions (exact match, contains, LLM-graded, similarity), and surfaces regressions. The red-teaming module generates adversarial inputs to test jailbreak resistance and safety guardrails.

Promptfoo targets AI engineers, product teams, and security researchers who need repeatable LLM testing. It replaces manual prompt testing with automated evaluation pipelines that run in CI/CD.

How it saves time or tokens

Manual prompt testing means running inputs one by one and eyeballing outputs. Promptfoo automates this with batch evaluation and structured scoring. You define test cases once and re-run them every time you change a prompt, model, or system configuration. The comparison view shows side-by-side results across models, making it obvious which performs better.

How to use

1. Install Promptfoo via npm.
2. Create a configuration file with prompts, providers, and test cases.
3. Run the eval and open the results viewer.

npm install -g promptfoo

# Initialize a config
promptfoo init

# Run evaluation
promptfoo eval

# Open results in browser
promptfoo view

Example

# promptfooconfig.yaml
prompts:
  - 'Summarize this article in 3 bullet points: {{article}}'

providers:
  - openai:gpt-4o
  - anthropic:claude-sonnet-4-20250514

tests:
  - vars:
      article: 'The Federal Reserve held interest rates steady...'
    assert:
      - type: contains
        value: 'interest rate'
      - type: llm-rubric
        value: 'Output should contain exactly 3 bullet points'
      - type: cost
        threshold: 0.01

Related on TokRepo

• AI tools for testing — Automated testing frameworks for AI applications
• Prompt library — Reusable prompts to test with Promptfoo

Common pitfalls

• LLM-graded assertions (llm-rubric) consume additional tokens for the grading call. Use deterministic assertions (contains, regex) where possible and reserve LLM grading for subjective quality checks.
• Running evaluations against many providers in parallel can hit rate limits. Configure concurrency limits in the config file.
• Red-team tests may produce harmful content in outputs. Run red-team evaluations in isolated environments and restrict access to the results.