Prompt Management

YuHeng features a decoupled prompt management system that keeps AI instructions separate from the application logic. This allows for easier iteration on AI behavior without changing code.

Architecture

The system is centered around the PromptManager class in lib/prompts.ts:

• Template Storage: Prompts are stored as .txt files in the prompts/ directory at the project root.
• Variable Injection: Uses a simple `` syntax for dynamic data injection.
• Singleton Pattern: The promptManager is a singleton instance used throughout the backend API.

How it Works

1. Prompt Definition

A typical prompt file (e.g., prompts/dish-init-prompt.txt) contains base instructions and placeholders for runtime data:

You are a professional nutritionist. 
Please analyze this food photo in .
Return the results in ...

2. Variable Replacement

When the code requests a prompt, it passes a variables object:

const prompt = await promptManager.getPrompt('dish-init-prompt', {
    language: 'English',
    unit: 'kcal'
});

The PromptManager reads the file and replaces all occurrences of and with the provided values.

Benefits

• Language Agnostic: Prompts can be easily adapted for multiple languages by injecting a language variable or using different prompt sets.
• Clear Separation: Nutritionists or prompt engineers can modify AI behavior by editing text files, without needing to understand the underlying TypeScript code.
• Version Control: Changes to AI behavior are tracked via Git as simple text diffs.