Skip to main content

System Prompts Overview

System prompts are the foundational instructions that guide GPT-5.1's behavior when generating gift recommendations. They define tone, format, language rules, and quality controls.

Purpose​

System prompts ensure:

  • Consistent tone across all responses
  • Proper formatting of product recommendations
  • Language accuracy (Estonian and English)
  • Quality controls to prevent hallucinations
  • Conversion optimization for better recommendations

Architecture​

Language Selection Rules​

Location: app/api/chat/system-prompt.ts:246-267

Rule 1: Estonian Query → Estonian Prompt​

if (language === 'et') {
  return buildEstonianSystemPrompt();
}

Rule 2: English Query → English Prompt​

if (language === 'en') {
  return buildEnglishSystemPrompt();
}

Rule 3: Mixed/Undefined → Estonian (Default)​

if (!language || language === 'mixed') {
  return buildEstonianSystemPrompt(); // Default to primary market
}

Prompt Components​

Each system prompt contains:

  1. Identity & Role

    • "I am RahvaRaamat AI shopping assistant"
    • Clear purpose statement

  2. Response Style

    • Tone configuration (enthusiastic/friendly/professional)
    • Sentence count per product (3-4)
    • Conversion language flags

  3. Formatting Rules

    • Numbered lists (1., 2., 3.)
    • Bold product titles
    • Structured descriptions

  4. Critical Prohibitions

    • No hallucinating content
    • No website URLs
    • No invented details

  5. Context Enrichment (optional)

    • When to ask questions
    • How many to ask
    • What to ask about

  6. Context Reminders (multi-turn)

    • Recipient context
    • Occasion context
    • Coherence across turns

Configuration-Driven​

Location: app/chat/config.ts

export const chatConfig = {
  productDescriptions: {
    tone: 'enthusiastic' | 'friendly' | 'professional',
    sentencesPerProduct: 3,
    useConversionLanguage: boolean,
    emphasizeBenefits: boolean,
    personalizeToContext: boolean
  },
  
  contextEnrichment: {
    enabled: boolean,
    maxQuestions: number,
    requireProducts: boolean,
    minConfidence: number
  }
};

Changes to config automatically update prompts - no code changes needed.

Dynamic Generation​

function generateDynamicSystemPrompt(
  language?: 'et' | 'en' | 'mixed',
  context?: { recipient?: string; occasion?: string }
): string {
  // 1. Select base prompt by language
  let basePrompt = getBaseSystemPrompt(language);
  
  // 2. Add context reminders if available
  if (context?.recipient) {
    basePrompt += contextReminder(context.recipient);
  }
  
  // 3. Add enrichment section if enabled
  if (chatConfig.contextEnrichment.enabled) {
    basePrompt += enrichmentInstructions();
  }
  
  return basePrompt;
}

Validation Pipeline​

After GPT-5.1 generates a response:

Key Features​

Multi-Language Support​

  • Native Estonian and English prompts
  • Cultural context awareness
  • Proper grammar and idioms

Hallucination Prevention​

  • Strict rules against content invention
  • Product-only descriptions
  • Validation checks post-generation

Conversion Optimization​

  • Enthusiastic tone options
  • Benefit-focused language
  • Call-to-action phrasing

Context Coherence​

  • Multi-turn conversation memory
  • Recipient/occasion reminders
  • Consistent personalization

File Structure​

app/api/chat/
├── system-prompt.ts           # Main prompt logic
├── config.ts                  # Configuration
└── utils/
    └── context-enrichment-questions.ts  # Question generation

Documentation Pages​

Performance Impact​

{
  promptBuildTime: ~5ms,        // Negligible
  promptTokens: ~400,           // Moderate context
  cachingEnabled: true,         // System prompt cached
  cacheHitRate: ~95%           // Very high reuse
}