Egged Agam Agent

​

Overview

​

Skills

​

Key Tools

​

Prompting Techniques

CRITICAL RESTRICTION :
- IF the data is not in the tools, state you cannot find it.
- DO NOT hallucinate or invent times/routes.

- Maintain the **female gender** in Hebrew (מדברת, בודקת, ממליצה) at all times.

**Audience Profile:** Elderly people, Haredi community, limited technical/Hebrew literacy.
**Tone Guidelines:**
- Respectful, slow-paced, patient, and very clear.
- Use short, simple sentences.
- **NO** technical jargon
- **NO** information overload: Give one clear answer at a time.

Step 1: Missing Information → Call missing_data_policy()
Step 2: Query Reconstruction
Step 3: Data Arrangement → Call field_judge()
Step 4: Action Execution
Step 5: Response Generation

​

Lessons Learned

• Explicit step-by-step workflows ensure consistent tool calling order
• Strict “no hallucination” rules with fallback responses maintain accuracy
• Audience-specific tone guidelines (elderly, Haredi) improve accessibility
• Language policy (Hebrew only) prevents confusion in multilingual scenarios
• Duration formatting rules (“x שעות ו-y דקות”) ensure consistent output

• Multiple similar tools (two line_details tools exist) - consolidated naming needed
• Balancing completeness vs. information overload for accessibility-focused audience
• Hebrew grammar maintenance (female gender) requires explicit instruction

​

Deep Dive

​

1. field_judge(query) - Query Parser & Intent Classifier

1. Time Context Calculation - Computes current Israel local time, date, and weekday in both Hebrew and English, accounting for DST
2. Relative Date Resolution - Converts Hebrew/English relative dates to concrete YYYY-MM-DD format:
   • “היום/today” → current date
   • “מחר/tomorrow” → +1 day
   • “מחרתיים” → +2 days
   • Weekday names like “יום ראשון” → next occurrence of that day
   • “שבוע הבא” (next week) handling
3. LLM Feature Extraction - Uses GPT with structured output to extract:

{
  "date": "YYYY-MM-DD",
  "fromTime": "HH:mm",
  "toTime": "HH:mm",
  "fromAddress": "string",
  "toAddress": "string",
  "fromStation": "string",
  "toStation": "string",
  "lineIdentifier": "string (e.g., '5א')",
  "transitType": "BUS|RAIL|TRAM|etc."
}

4. Intent Classification - Parallel LLM call to classify as:
   • LINE_SCHEDULE - Questions about specific bus lines, schedules, first/last departures
   • DIRECTIONS - Route planning from A to B
5. Station Name Refinement - For fuzzy station names, calls Moovit API to get candidate stops, then uses LLM to pick the best match
6. Persistence - Saves extracted features and intent to context.kv for downstream tools

• Extensive rules for Hebrew date/time parsing
• Special handling for “מוצאי שבת” (Saturday night)
• Address formatting rules (comma-separated, not “in”)
• Station number priority over station name
• Line identifier preservation (keeps Hebrew suffixes like “37א”)

​

2. missing_data_policy(query) - Smart Data Validation - Mini Ai Agent

1. Street Length Check Tool - Defines an internal tool that the internal LLM uses when needed (llm choice) that uses Google Maps API to:
   • Geocode street names
   • Calculate street length from viewport bounds
   • If street > 0.7km, prompts for building number (streets in Israel can be long)
2. LLM Analysis - Uses GPT to analyze the query against a comprehensive ruleset:
   • Universal Blockers:
     • AM/PM ambiguity (e.g., “at 5” needs clarification)
     • Station number overrides (no follow-up needed)
     • One question per piece of info
   • Route Planning Requirements:
     • Origin city + Destination city + Time
     • Street name recognition without “רחוב” prefix
     • City-to-city queries are sufficient (no intra-city details needed)
   • Line Schedule Requirements:
     • Line number OR location is sufficient
     • Station name without city → ask for city
3. Question Generation - Returns JSON array of English questions to ask the user
4. System Message Delivery - Sends questions via context.agent.sendSystemMessage()

// Street length check determines if building number needed
if (lengthKm > 0.7) {
  // Street is long → ask for building number
} else {
  // Street is short → building number not needed
}

​

3. transportation_data() - Moovit API Gateway

​

A. runLineSchedule() - For LINE_SCHEDULE intent

1. API Call - POST to Moovit’s SearchAI endpoint with:

{
  "intent": "LINE_SCHEDULE",
  "userLanguage": "he",
  "searchLocation": { "location": { "latitude": 31917827, "longitude": 34803918 } },
  "features": { /* from KV */ }
}

2. Response Processing:
   • Filters and sorts lines (Egged first)
   • Limits to 5 lines max
   • Normalizes stop names (e.g., “ת.” → “תחנה ”)
   • Transliterates line number suffixes (“37א” → “37 alef” for TTS)
   • Detects duplicate lines (same line in opposite directions)
   • Identifies live arrival data availability
3. Output:

{
  "linesSchedule": { "date": "...", "lines": [...] },
  "error": "",
  "current_time": "2025-01-15 14:30"
}

​

B. runDirections() - For DIRECTIONS intent

1. API Call - Same endpoint but with:

{ "intent": "DIRECTIONS", ... }

2. Response Processing:
   • Extracts itinerary with legs
   • Processes each leg (Bus, Walk, Wait, etc.)
   • Flattens location names
   • Removes Wait/Walk/PathWayWalk legs from final output
   • Calculates total trip time from startTime/endTime
   • Sorts legs by startTime
3. Output:

{
  "itinerary": { "legs": [...], "total_trip_time": "1h 30m" },
  "error": "",
  "current_time": "2025-01-15 14:30"
}

• Retry logic (3 attempts with 100ms delay)
• 424 status special handling (surface error text)
• Hebrew stop name normalization
• Metadata persistence for debugging
• Current time in Israel timezone

​

How They Work Together

User Query: "מתי יוצא קו 5 מחיפה מחר בבוקר?"

1. missing_data_policy() → Checks if all required info present
   ↓ (no questions needed)
   
2. field_judge() → Extracts:
   - lineIdentifier: "5"
   - fromAddress: "חיפה"
   - date: "2025-01-16"
   - fromTime: "09:00"
   - intent: "LINE_SCHEDULE"
   ↓ (saves to KV)
   
3. transportation_data() → Reads intent from KV
   ↓ Calls Moovit LINE_SCHEDULE API
   ↓ Returns schedule data
   
4. Agent formats response based on prompt rules