PromptSearch

Automated Reverse Prompt Engineering: Evolve system prompts from target outputs.

Installation

pip install promptsearch

Quick Start

from promptsearch import PromptSearcher

# Define your target output (what you want the LLM to produce)
target_output = {
    "phone": "555-1234",
    "name": "John Doe"
}

# Define your initial system prompt
initial_prompt = "Extract information from the text."

# Create the searcher
searcher = PromptSearcher(
    target_output=target_output,
    initial_prompt=initial_prompt
)

# Run optimization
train_input = "Contact John Doe at 555-1234 for more information."
result = searcher.optimize(train_input=train_input, generations=5)

print(f"Best Prompt: {result['best_prompt']}")
print(f"Best Score: {result['best_score']:.3f}")
print(f"Best Output: {result['best_output']}")

How It Works

PromptSearch uses evolutionary hill climbing to optimize system prompts:

1. Generate: Test the current prompt with your input
2. Score: Compare output to target using semantic similarity
3. Mutate: Use an LLM to rewrite the prompt based on failure analysis
4. Repeat: Continue until convergence or max generations

Requirements

• Python 3.8+
• OpenAI API key (set OPENAI_API_KEY environment variable)

License

MIT License - see LICENSE file for details.