docs/features/word-problems.md

Word Problem Handling Strategy

Overview

meeTARA Agent handles word problems differently based on their complexity and type:

Three Types of Word Problems

1. Simple Math Word Problems → Calculator ✅

Examples:

• "What's 15% of 340?" → Extracted as 0.15 * 340 → Calculator calculates
• "Find 25 * 48" → Calculator handles directly
• "Calculate tan(45)" → Scientific function → Calculator

Why Calculator:

• Clear mathematical expression can be extracted via pattern matching
• Calculator ensures 100% accuracy
• Fast and efficient

2. Complex Word Problems → Model Directly ✅

Examples:

• "If John has 5 apples and gives 2 away, how many does he have left?"
• "A train travels 60 miles in 2 hours. What's its speed?"
• "Sarah buys 3 books at $10 each and 2 pens at $5 each. How much did she spend?"
• "If a rectangle has length 6 cm and width 4 cm, what's its area?" (geometry - has keywords)

Why Model Directly:

• Requires understanding relationships between entities (John, apples, gives away)
• Need to extract formula from context (speed = distance/time)
• Need to combine multiple operations (3 books × $10 + 2 pens × $5)
• Model can understand natural language relationships and extract math
• Calculator would need perfect extraction, which is hard for complex relationships

Detection:

• Uses word_problem_keywords from config: "if", "has", "gives", "travels", "buys", "costs", etc.
• Strategy: model_direct (configurable in agent_config.json)
• Skips calculator if complex word problem detected → Sends directly to model

3. Word Problems About Current Events/Data → Web Search + Model ✅

Examples:

• "What's the current unemployment rate in the US?"
• "What's the current price of Bitcoin?"
• "What's today's stock market performance?"

Why Web Search:

• Requires up-to-date information beyond training data
• Real-time data that changes frequently
• Web search provides current data → Model structures response

Configuration

All word problem handling is configurable in config/agent_config.json:

{
  "calculator": {
    "word_problem_keywords": [
      "if", "when", "has", "gives", "receives", "buys", "sells",
      "travels", "drives", "costs", "price", "total", "spends",
      "left", "remaining", "more than", "less than", "times"
    ],
    "word_problem_strategy": "model_direct",
    "word_problem_explanation": "Simple word problems handled by calculator. Complex word problems go to model. Current events go to web search."
  }
}

Decision Flow

User Query: Word Problem
    ↓
Check if contains word_problem_keywords (if, has, gives, etc.)
    ↓
┌─────────────────────────────────────────────────┐
│ Is it about current events/data?                │
│ → Contains "current", "today", "latest", etc.   │
└─────────────────────────────────────────────────┘
    YES → Web Search → Model
    NO  ↓
┌─────────────────────────────────────────────────┐
│ Can math be extracted via patterns?             │
│ → "15% of 340" → 0.15 * 340 ✓                  │
│ → "tan(45)" → tan(45) ✓                        │
│ → "25 * 48" → 25 * 48 ✓                        │
└─────────────────────────────────────────────────┘
    YES → Calculator → Model
    NO  ↓
┌─────────────────────────────────────────────────┐
│ Is it a complex word problem?                   │
│ → Contains relationship keywords                │
│ → No clear math expression extractable          │
└─────────────────────────────────────────────────┘
    YES → Model Directly (skip calculator)
    NO  → Model Directly (fallback)

Examples

Example 1: Simple → Calculator

Query: "What's 15% of 340?"

• Detection: Percentage pattern matched
• Extraction: 0.15 * 340
• Calculator: 51
• Model: Structures response with steps

Example 2: Complex → Model Directly

Query: "If John has 5 apples and gives 2 away, how many does he have left?"

• Detection: Contains word_problem_keywords ("if", "has", "gives", "left")
• Complex word problem: Yes (requires understanding relationships)
• Calculator: Skipped (no clear expression to extract)
• Model: Understands relationship → Extracts: 5 - 2 = 3 → Responds

Example 3: Current Events → Web Search

Query: "What's the current unemployment rate?"

• Detection: Contains "current" → Web search keyword
• Web Search: Fetches current data
• Model: Structures response with current data

Example 4: Geometry → Full Query to Model

Query: "Find the surface area of a rectangular prism with dimensions: length = 6 cm, width = 4 cm, height = 5 cm"

• Detection: Geometry keywords ("surface area", "rectangular prism")
• Calculator: Receives full query (not just numbers)
• Calculator: Detects geometry → Returns context message
• Model: Calculates using formula 2(lw + lh + wh) = 148 cm²

Recommendations

1. Simple Math → Calculator (100% accurate, fast)
2. Complex Word Problems → Model (understands relationships, extracts math)
3. Current Events → Web Search (up-to-date data)
4. Geometry → Full query to calculator/model (needs formula context)

Why This Strategy Works

• Calculator = Perfect for clear math expressions (ensures accuracy)
• Model = Perfect for understanding relationships and extracting formulas
• Web Search = Perfect for current/realtime data
• Combination = Best of all worlds (tool precision + model understanding + current data)