PAL - Prompt Assembly Language

PAL (Prompt Assembly Language) is a framework for managing LLM prompts as versioned, composable software artifacts. It treats prompt engineering with the same rigor as software engineering, focusing on modularity, versioning, and testability.

See also the NodeJS version of PAL (WIP)

⚡ Features

• Modular Components: Break prompts into reusable, versioned components
• Template System: Powerful Jinja2-based templating with variable injection
• Dependency Management: Import and compose components from local files or URLs
• LLM Integration: Built-in support for OpenAI, Anthropic, and custom providers
• Evaluation Framework: Comprehensive testing system for prompt validation
• Rich CLI: Beautiful command-line interface with syntax highlighting
• Flexible Extensions: Use .pal/.pal.lib or .yml/.lib.yml extensions
• Type Safety: Full Pydantic v2 validation for all schemas
• Observability: Structured logging and execution tracking

📦 Installation

# Install with uv (recommended)
uv add pal-framework

# Or with pip
pip install pal-framework

📁 Project Structure

my_pal_project/
├── prompts/
│   ├── classify_intent.pal     # or .yml for better IDE support
│   └── code_review.pal
├── libraries/
│   ├── behavioral_traits.pal.lib    # or .lib.yml
│   ├── reasoning_strategies.pal.lib
│   └── output_formats.pal.lib
└── evaluation/
    └── classify_intent.eval.yaml

🚀 Quick Start

1. Create a Component Library

For a detailed guide, read this.

# libraries/traits.pal.lib
pal_version: "1.0"
library_id: "com.example.traits"
version: "1.0.0"
description: "Behavioral traits for AI agents"
type: "trait"

components:
  - name: "helpful_assistant"
    description: "A helpful and polite assistant"
    content: |
      You are a helpful, harmless, and honest AI assistant. You provide
      accurate information while being respectful and considerate.

2. Create a Prompt Assembly

For a detailed guide, read this.

# prompts/classify_intent.pal
pal_version: "1.0"
id: "classify-user-intent"
version: "1.0.0"
description: "Classifies user queries into intent categories"

imports:
  traits: "./libraries/traits.pal.lib"

variables:
  - name: "user_query"
    type: "string"
    description: "The user's input query"
  - name: "available_intents"
    type: "list"
    description: "List of available intent categories"

composition:
  - "{{ traits.helpful_assistant }}"
  - |
    ## Task
    Classify this user query into one of the available intents:

    **Available Intents:**
    {% for intent in available_intents %}
    - {{ intent.name }}: {{ intent.description }}
    {% endfor %}

    **User Query:** {{ user_query }}

3. Use the CLI

# Compile a prompt
pal compile prompts/classify_intent.pal --vars '{"user_query": "Take me to google.com", "available_intents": [{"name": "navigate", "description": "Go to URL"}]}'

# Execute with an LLM
pal execute prompts/classify_intent.pal --model gpt-4 --provider openai --vars '{"user_query": "Take me to google.com", "available_intents": [{"name": "navigate", "description": "Go to URL"}]}'

# Validate PAL files
pal validate prompts/ --recursive

# Run evaluation tests
pal evaluate evaluation/classify_intent.eval.yaml

4. Use Programmatically

import asyncio
from pal import PromptCompiler, PromptExecutor, MockLLMClient

async def main():
    # Set up components
    compiler = PromptCompiler()
    llm_client = MockLLMClient("Mock response")
    executor = PromptExecutor(llm_client)

    # Compile prompt
    variables = {
        "user_query": "What's the weather?",
        "available_intents": [{"name": "search", "description": "Search for info"}]
    }

    compiled_prompt = await compiler.compile_from_file(
        "prompts/classify_intent.pal",
        variables
    )

    print("Compiled Prompt:", compiled_prompt)

asyncio.run(main())

🧪 Evaluation System

Create test suites to validate your prompts:

# evaluation/classify_intent.eval.yaml
pal_version: "1.0"
prompt_id: "classify-user-intent"
target_version: "1.0.0"

test_cases:
  - name: "navigation_test"
    variables:
      user_query: "Go to google.com"
      available_intents: [{ "name": "navigate", "description": "Visit URL" }]
    assertions:
      - type: "json_valid"
      - type: "contains"
        config:
          text: "navigate"

🏗️ Architecture

PAL follows modern software engineering principles:

• Schema Validation: All files are validated against strict Pydantic schemas
• Dependency Resolution: Automatic import resolution with circular dependency detection
• Template Engine: Jinja2 for powerful variable interpolation and logic
• Observability: Structured logging with execution metrics and cost tracking
• Type Safety: Full type hints and runtime validation

🛠️ CLI Commands

🧩 Component Types

PAL supports different types of reusable components:

• persona: AI personality and role definitions
• task: Specific instructions or objectives
• context: Background information and knowledge
• rules: Constraints and guidelines
• examples: Few-shot learning examples
• output_schema: Output format specifications
• reasoning: Thinking strategies and methodologies
• trait: Behavioral characteristics
• note: Documentation and comments

🤝 Contributing

We welcome contributions! Please see our Contributing Guide for details.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🆘 Support

• 📚 Documentation
• 🐛 Issues
• 💬 Discussions

🗺️ Roadmap

• PAL Registry: Centralized repository for sharing components
• Visual Builder: Drag-and-drop prompt composition interface
• IDE Extensions: VS Code and other editor integrations