Progressive Context Building

The system builds understanding progressively through clarifying questions when initial context is vague.

Multi-Turn Clarification Flow

Progressive State

Turn	Context State	Confidence	Action
1	{ intent: "unknown" }	0.3	Ask: "Kellele?"
2	{ recipient: "kolleeg" }	0.6	Ask: "Mis puhuks?"
3	{ recipient: "kolleeg", occasion: "sünnipäev" }	0.8	Search

Context Accumulation Logic

export async function orchestrate(options): Promise<Result> {
  // Step 1: Extract from current message
  const { giftContext } = await extractContext({
    userMessage,
    conversationState  // Prior state
  });
  
  // Step 2: Fetch stored context
  const { storedContext } = await fetchStoredContext({
    conversationId,
    giftContext
  });
  
  // Step 3: Merge stored with new
  if (!giftContext.budget && storedContext?.budget) {
    giftContext.budget = storedContext.budget;  // Restore
  }
  
  if (!giftContext.occasion && storedContext?.occasion) {
    giftContext.occasion = storedContext.occasion;  // Restore
  }
  
  // ... similar for recipient, productType, category
  
  return { giftContext, storedContext };
}

When to Ask Clarifying Questions

export function shouldAskClarifyingQuestion(
  giftContext: GiftContext
): boolean {
  // Low confidence OR ambiguous intent
  const hasLowConfidence = (giftContext.confidence ?? 0) < 0.7;
  const isAmbiguous = ['unknown', 'greeting'].includes(giftContext.intent);
  
  if (!hasLowConfidence && !isAmbiguous) return false;
  
  // Has meaningful signals?
  const hasProductSignals = hasMeaningfulProductSignals(giftContext);
  const hasGiftContext = !isGiftContextMissing(giftContext);
  
  // Has explicit productType/category?
  const hasExplicit = giftContext.productType || giftContext.category;
  
  // ALWAYS search if explicit present
  if (hasExplicit) return false;
  
  // Ask if vague AND no signals AND no context
  return (hasLowConfidence || isAmbiguous) && 
         !hasProductSignals && 
         !hasGiftContext;
}

Clarifying Question Generation

System Prompt Enhancement:

prompt += `
CONTEXT ENRICHMENT:
After recommendations, ask up to 2 clarifying questions:

Ask ONLY about missing:
- Budget unknown → "What's your budget?"
- Recipient unknown → "Who is this gift for?"
- Occasion unknown → "What's the occasion?"
- Age unknown → "How old is [recipient]?"
- Interests unknown → "What are [recipient]'s interests?"

Format naturally:
- "To help me give better recommendations:" + questions

Rules:
- Maximum 2 questions
- DON'T ask if already known
- Natural and conversational
`;

Real Example

Turn 1:

User: "Kingitust"
System: "Kellele soovite kingitust valida? Sõbrale, perele, või kolleegile?

Samuti aitaks teada, mis puhuks:
- Sünnipäev
- Jõulud  
- Tänuavaldus

Kui teil on eelarve piirang, andke teada!"

Progressive Metrics

Metric	Value	Notes
Avg Turns to Specificity	2.3	Vague → Specific
Clarification Success Rate	89%	User provides context
Abandonment After Clarification	11%	User gives up
Context Accumulation Accuracy	94%	Fields correctly merged

Related Documentation

• Refinement Detection - How refinements are detected
• Budget System - Budget handling details
• Overview - Conversational intelligence overview