promptfoo-evaluation

Document Skill original

Rendu depuis le dépôt source en conservant titres, exemples, code, tableaux, liens et images.

Promptfoo Evaluation

Overview

This skill provides guidance for configuring and running LLM evaluations using Promptfoo, an open-source CLI tool for testing and comparing LLM outputs.

Quick Start

bash

# Initialize a new evaluation project
npx promptfoo@latest init

# Run evaluation
npx promptfoo@latest eval

# View results in browser
npx promptfoo@latest view

Configuration Structure

A typical Promptfoo project structure:

project/
├── promptfooconfig.yaml    # Main configuration
├── prompts/
│   ├── system.md           # System prompt
│   └── chat.json           # Chat format prompt
├── tests/
│   └── cases.yaml          # Test cases
└── scripts/
    └── metrics.py          # Custom Python assertions

Core Configuration (promptfooconfig.yaml)

yaml

# yaml-language-server: $schema=https://promptfoo.dev/config-schema.json
description: "My LLM Evaluation"

# Prompts to test
prompts:
  - file://prompts/system.md
  - file://prompts/chat.json

# Models to compare
providers:
  - id: anthropic:messages:claude-sonnet-4-6
    label: Claude-Sonnet-4.6
  - id: openai:gpt-4.1
    label: GPT-4.1

# Test cases
tests: file://tests/cases.yaml

# Concurrency control (MUST be under commandLineOptions, NOT top-level)
commandLineOptions:
  maxConcurrency: 2

# Default assertions for all tests
defaultTest:
  assert:
    - type: python
      value: file://scripts/metrics.py:custom_assert
    - type: llm-rubric
      value: |
        Evaluate the response quality on a 0-1 scale.
      threshold: 0.7

# Output path
outputPath: results/eval-results.json

Prompt Formats

Text Prompt (system.md)

markdown

You are a helpful assistant.

Task: {{task}}
Context: {{context}}

Chat Format (chat.json)

json

[
  {"role": "system", "content": "{{system_prompt}}"},
  {"role": "user", "content": "{{user_input}}"}
]

Few-Shot Pattern

Embed examples directly in prompt or use chat format with assistant messages:

json

[
  {"role": "system", "content": "{{system_prompt}}"},
  {"role": "user", "content": "Example input: {{example_input}}"},
  {"role": "assistant", "content": "{{example_output}}"},
  {"role": "user", "content": "Now process: {{actual_input}}"}
]

Test Cases (tests/cases.yaml)

yaml

- description: "Test case 1"
  vars:
    system_prompt: file://prompts/system.md
    user_input: "Hello world"
    # Load content from files
    context: file://data/context.txt
  assert:
    - type: contains
      value: "expected text"
    - type: python
      value: file://scripts/metrics.py:custom_check
      threshold: 0.8

Python Custom Assertions

Create a Python file for custom assertions (e.g., scripts/metrics.py):

python

def get_assert(output: str, context: dict) -> dict:
    """Default assertion function."""
    vars_dict = context.get('vars', {})

    # Access test variables
    expected = vars_dict.get('expected', '')

    # Return result
    return {
        "pass": expected in output,
        "score": 0.8,
        "reason": "Contains expected content",
        "named_scores": {"relevance": 0.9}
    }

def custom_check(output: str, context: dict) -> dict:
    """Custom named assertion."""
    word_count = len(output.split())
    passed = 100 <= word_count <= 500

    return {
        "pass": passed,
        "score": min(1.0, word_count / 300),
        "reason": f"Word count: {word_count}"
    }

Key points:

Default function name is get_assert
Specify function with file://path.py:function_name
Return bool, float (score), or dict with pass/score/reason
Access variables via context['vars']

LLM-as-Judge (llm-rubric)

yaml

assert:
  - type: llm-rubric
    value: |
      Evaluate the response based on:
      1. Accuracy of information
      2. Clarity of explanation
      3. Completeness

      Score 0.0-1.0 where 0.7+ is passing.
    threshold: 0.7
    provider: openai:gpt-4.1  # Optional: override grader model

When using a relay/proxy API, each llm-rubric assertion needs its own provider config with apiBaseUrl. Otherwise the grader falls back to the default Anthropic/OpenAI endpoint and gets 401 errors:

yaml

assert:
  - type: llm-rubric
    value: |
      Evaluate quality on a 0-1 scale.
    threshold: 0.7
    provider:
      id: anthropic:messages:claude-sonnet-4-6
      config:
        apiBaseUrl: https://your-relay.example.com/api

Best practices:

Provide clear scoring criteria
Use threshold to set minimum passing score
Default grader uses available API keys (OpenAI → Anthropic → Google)
When using relay/proxy: every llm-rubric must have its own provider with apiBaseUrl — the main provider's apiBaseUrl is NOT inherited

Common Assertion Types

Type	Usage	Example
`contains`	Check substring	`value: "hello"`
`icontains`	Case-insensitive	`value: "HELLO"`
`equals`	Exact match	`value: "42"`
`regex`	Pattern match	`value: "\\d{4}"`
`python`	Custom logic	`value: file://script.py`
`llm-rubric`	LLM grading	`value: "Is professional"`
`latency`	Response time	`threshold: 1000`

File References

All file:// paths are resolved relative to promptfooconfig.yaml location (NOT the YAML file containing the reference). This is a common gotcha when tests: references a separate YAML file — the file:// paths inside that test file still resolve from the config root.

yaml

# Load file content as variable
vars:
  content: file://data/input.txt

# Load prompt from file
prompts:
  - file://prompts/main.md

# Load test cases from file
tests: file://tests/cases.yaml

# Load Python assertion
assert:
  - type: python
    value: file://scripts/check.py:validate

Running Evaluations

bash

# Basic run
npx promptfoo@latest eval

# With specific config
npx promptfoo@latest eval --config path/to/config.yaml

# Output to file
npx promptfoo@latest eval --output results.json

# Filter tests
npx promptfoo@latest eval --filter-metadata category=math

# View results
npx promptfoo@latest view

Relay / Proxy API Configuration

When using an API relay or proxy instead of direct Anthropic/OpenAI endpoints:

yaml

providers:
  - id: anthropic:messages:claude-sonnet-4-6
    label: Claude-Sonnet-4.6
    config:
      max_tokens: 4096
      apiBaseUrl: https://your-relay.example.com/api  # Promptfoo appends /v1/messages

# CRITICAL: maxConcurrency MUST be under commandLineOptions (NOT top-level)
commandLineOptions:
  maxConcurrency: 1  # Respect relay rate limits

Key rules:

apiBaseUrl goes in providers[].config — Promptfoo appends /v1/messages automatically
maxConcurrency must be under commandLineOptions: — placing it at top level is silently ignored
When using relay with LLM-as-judge, set maxConcurrency: 1 to avoid concurrent request limits (generation + grading share the same pool)
Pass relay token as ANTHROPIC_API_KEY env var

Troubleshooting

Python not found:

bash

export PROMPTFOO_PYTHON=python3

Large outputs truncated: Outputs over 30000 characters are truncated. Use head_limit in assertions.

File not found errors: All file:// paths resolve relative to promptfooconfig.yaml location.

maxConcurrency ignored (shows "up to N at a time"): maxConcurrency must be under commandLineOptions:, not at the YAML top level. This is a common mistake.

LLM-as-judge returns 401 with relay API: Each llm-rubric assertion must have its own provider with apiBaseUrl. The main provider config is not inherited by grader assertions.

HTML tags in model output inflating metrics: Models may output <br>, <b>, etc. in structured content. Strip HTML in Python assertions before measuring:

python

import re
clean_text = re.sub(r'<[^>]+>', '', raw_text)

Echo Provider (Preview Mode)

Use the echo provider to preview rendered prompts without making API calls:

yaml

# promptfooconfig-preview.yaml
providers:
  - echo  # Returns prompt as output, no API calls

tests:
  - vars:
      input: "test content"

Use cases:

Preview prompt rendering before expensive API calls
Verify Few-shot examples are loaded correctly
Debug variable substitution issues
Validate prompt structure

bash

# Run preview mode
npx promptfoo@latest eval --config promptfooconfig-preview.yaml

Cost: Free - no API tokens consumed.

Advanced Few-Shot Implementation

Multi-turn Conversation Pattern

For complex few-shot learning with full examples:

json

[
  {"role": "system", "content": "{{system_prompt}}"},

  // Few-shot Example 1
  {"role": "user", "content": "Task: {{example_input_1}}"},
  {"role": "assistant", "content": "{{example_output_1}}"},

  // Few-shot Example 2 (optional)
  {"role": "user", "content": "Task: {{example_input_2}}"},
  {"role": "assistant", "content": "{{example_output_2}}"},

  // Actual test
  {"role": "user", "content": "Task: {{actual_input}}"}
]

Test case configuration:

yaml

tests:
  - vars:
      system_prompt: file://prompts/system.md
      # Few-shot examples
      example_input_1: file://data/examples/input1.txt
      example_output_1: file://data/examples/output1.txt
      example_input_2: file://data/examples/input2.txt
      example_output_2: file://data/examples/output2.txt
      # Actual test
      actual_input: file://data/test1.txt

Best practices:

Use 1-3 few-shot examples (more may dilute effectiveness)
Ensure examples match the task format exactly
Load examples from files for better maintainability
Use echo provider first to verify structure

Long Text Handling

For Chinese/long-form content evaluations (10k+ characters):

Configuration:

yaml

providers:
  - id: anthropic:messages:claude-sonnet-4-6
    config:
      max_tokens: 8192  # Increase for long outputs

defaultTest:
  assert:
    - type: python
      value: file://scripts/metrics.py:check_length

Python assertion for text metrics:

python

import re

def strip_tags(text: str) -> str:
    """Remove HTML tags for pure text."""
    return re.sub(r'<[^>]+>', '', text)

def check_length(output: str, context: dict) -> dict:
    """Check output length constraints."""
    raw_input = context['vars'].get('raw_input', '')

    input_len = len(strip_tags(raw_input))
    output_len = len(strip_tags(output))

    reduction_ratio = 1 - (output_len / input_len) if input_len > 0 else 0

    return {
        "pass": 0.7 <= reduction_ratio <= 0.9,
        "score": reduction_ratio,
        "reason": f"Reduction: {reduction_ratio:.1%} (target: 70-90%)",
        "named_scores": {
            "input_length": input_len,
            "output_length": output_len,
            "reduction_ratio": reduction_ratio
        }
    }

Real-World Example

Project: Chinese short-video content curation from long transcripts

Structure:

tiaogaoren/
├── promptfooconfig.yaml          # Production config
├── promptfooconfig-preview.yaml  # Preview config (echo provider)
├── prompts/
│   ├── tiaogaoren-prompt.json   # Chat format with few-shot
│   └── v4/system-v4.md          # System prompt
├── tests/cases.yaml              # 3 test samples
├── scripts/metrics.py            # Custom metrics (reduction ratio, etc.)
├── data/                         # 5 samples (2 few-shot, 3 eval)
└── results/

See: ./tiaogaoren/ (example project root) for full implementation.

Resources

For detailed API reference and advanced patterns, see references/promptfoo_api.md.