Priompt is a JSX-based library for designing prompts for Large Language Models (LLMs), enabling structured and prioritized content inclusion within limited context windows. It targets developers and researchers building complex LLM applications, offering a declarative approach to prompt engineering that mirrors web development patterns.

How It Works

Priompt renders prompts from JSX components, assigning priorities to elements to manage context window usage. It employs a binary search mechanism to find an optimal cutoff priority, ensuring the rendered prompt adheres to token limits while maximizing the inclusion of high-priority content. Key components like <scope>, <first>, and <empty> allow fine-grained control over content inclusion, fallback strategies, and token reservation.

Quick Start & Requirements

• Install: npm install @anysphere/priompt && npm install -D @anysphere/priompt-preview (or yarn/pnpm equivalents).
• Prerequisites: Node.js.
• Examples: Refer to examples/README.md.

Highlighted Details

• JSX-based declarative prompt authoring.
• Priority-based content inclusion for context window management.
• Components for messages (<UserMessage>, <AssistantMessage>, <SystemMessage>), tool definitions (<Tools>, <ZTools>), and more.
• Features like onEject/onInclude callbacks and sourcemaps for debugging.

Maintenance & Community

• MIT Licensed. Contributions are welcome.

Licensing & Compatibility

• MIT License. Permissive for commercial use and closed-source linking.

Limitations & Caveats

The library acknowledges that overusing priorities might lead to hard-to-cache prompts, potentially increasing LLM inference costs or latency. The current renderer has performance limitations with very large numbers of scopes (over 10K) and may not always guarantee the absolute optimal token cutoff due to potential edge cases in its binary search implementation.