promptfoo

Promptfoo and Vertex AI

Promptfoo is an open-source tool for evaluating and red-teaming (security) LLM apps. It's similar to DeepEval but with added focus on security.

In this tutorial, you'll learn how to use PromptFoo with Vertex AI.

Setup

Install PromptFoo in your environment. For example, on Mac OS, you can use brew:

brew install promptfoo

Verify that it's installed:

promptfoo --version

Make sure your gcloud is setup with your project:

gcloud config set core/project your-project-id

And you're logged in:

gcloud auth application-default login

You can start using Promptfoo in different scenarios with its interactive guide:

promptfoo init

? What would you like to do?
  Not sure yet
❯ Improve prompt and model performance
  Improve RAG performance
  Improve agent/chain of thought performance
  Run a red team evaluation

Evaluation

In Improve prompt and model performance, you can use Promptfoo to evaluate against different LLMs from OpenAI, Anthropic, Gemini, or simply an HTTP endpoint.

promptfooconfig1.yaml is a sample configuration for evaluating against a couple of Vertex AI models.

Run:

promptfoo eval -c promptfooconfig1.yaml

View the results in the console:

┌──────────────────────────────────────────────────────────────┬──────────────────────────────────────────────────────────────┬──────────────────────────────────────────────────────────────┐
│ question                                                     │ [vertex:gemini-1.5-flash-002] You are a helpful assistant.   │ [vertex:gemini-1.5-pro-002] You are a helpful assistant.     │
│                                                              │ Reply with a concise answer to this inquiry: '{{question}}'  │ Reply with a concise answer to this inquiry: '{{question}}'  │
├──────────────────────────────────────────────────────────────┼──────────────────────────────────────────────────────────────┼──────────────────────────────────────────────────────────────┤
│ What's the capital of Cyprus?                                │ [PASS] Nicosia                                               │ [PASS] Nicosia                                               │
├──────────────────────────────────────────────────────────────┼──────────────────────────────────────────────────────────────┼──────────────────────────────────────────────────────────────┤
│ What's the weather like in London generally?                 │ [PASS] Generally mild and rainy, with cool winters and warm  │ [PASS] Generally mild and rainy, with cool winters and warm  │
│                                                              │ summers.                                                     │ summers.                                                     │
└──────────────────────────────────────────────────────────────┴──────────────────────────────────────────────────────────────┴──────────────────────────────────────────────────────────────┘
==========================================================================================================================================================================================
✔ Evaluation complete.

» Run promptfoo view to use the local web viewer
» Run promptfoo share to create a shareable URL
» This project needs your feedback. What's one thing we can improve? https://forms.gle/YFLgTe1dKJKNSCsU7
==========================================================================================================================================================================================

Also view it in the browser:

promptfoo view

Red-teaming (security)

In Run a read team evaluation, you can use Promptfoo to do read-team security testing to find vulnerabilities by simulating malicious inputs against LLMs or your endpoints using LLMs.

promptfooconfig2.yaml is a sample configuration for read team testing against a Vertex AI model.

Generate test cases from this configuration:

promptfoo redteam generate -c promptfooconfig2.yaml

This generates a readteam.yaml file with the actual test cases:

Test Generation Summary:
• Total tests: 18
• Plugin tests: 6
• Plugins: 6
• Strategies: 2
• Max concurrency: 1

Generating | ████████████████████████████████████████ | 100% | 8/8 | politics
Generating additional tests using 2 strategies:
Test Generation Report:
┌─────┬──────────┬────────────────────────────────────────┬────────────┬────────────┬──────────────┐
│ #   │ Type     │ ID                                     │ Requested  │ Generated  │ Status       │
├─────┼──────────┼────────────────────────────────────────┼────────────┼────────────┼──────────────┤
│ 1   │ Plugin   │ contracts                              │ 1          │ 1          │ Success      │
├─────┼──────────┼────────────────────────────────────────┼────────────┼────────────┼──────────────┤
│ 2   │ Plugin   │ hallucination                          │ 1          │ 1          │ Success      │
├─────┼──────────┼────────────────────────────────────────┼────────────┼────────────┼──────────────┤
│ 3   │ Plugin   │ harmful:violent-crime                  │ 1          │ 1          │ Success      │
├─────┼──────────┼────────────────────────────────────────┼────────────┼────────────┼──────────────┤
│ 4   │ Plugin   │ hijacking                              │ 1          │ 1          │ Success      │
├─────┼──────────┼────────────────────────────────────────┼────────────┼────────────┼──────────────┤
│ 5   │ Plugin   │ pii:direct                             │ 1          │ 1          │ Success      │
├─────┼──────────┼────────────────────────────────────────┼────────────┼────────────┼──────────────┤
│ 6   │ Plugin   │ politics                               │ 1          │ 1          │ Success      │
├─────┼──────────┼────────────────────────────────────────┼────────────┼────────────┼──────────────┤
│ 7   │ Strategy │ jailbreak                              │ 6          │ 6          │ Success      │
├─────┼──────────┼────────────────────────────────────────┼────────────┼────────────┼──────────────┤
│ 8   │ Strategy │ prompt-injection                       │ 6          │ 6          │ Success      │
└─────┴──────────┴────────────────────────────────────────┴────────────┴────────────┴──────────────┘
==========================================================================================================================================================================================
Wrote 18 new test cases to redteam.yaml

Run the test cases:

promptfoo redteam eval

View your results:

References

• Google Vertex provider on PromptFoo
• Getting started
• LLM red teaming
• Promptfoo examples