Skip to main content
TripLoom’s AI copilot provides real-time, context-aware guidance across every stage of your trip. From initial planning to day-by-day itinerary support, the AI understands your trip constraints and delivers practical next steps.

​
How the AI works

The AI copilot is read-only and context-aware. It analyzes your trip data (destination, dates, travelers, budget) and the current page you’re viewing to provide targeted recommendations. It never modifies your bookings or expenses—only you can make those changes.
The AI assistant is available on every trip page. Open the AI panel from the bottom-right corner to start a conversation.

​
Page-aware assistance

The copilot adapts its guidance based on which page you’re viewing:
  • Suggests the highest-impact next action based on trip progress
  • Identifies missing essentials (flights, hotels, budget)
  • Recommends prioritized checklist items
  • Provides live flight status when you share a flight number and date (e.g., “AC123 2025-06-15”)
  • Compares price-time tradeoffs for different routes
  • Flags tight connection risks
  • Fetches bus and rail options when you use the format “from Berlin to Munich”
  • Compares reliability, duration, and transfers
  • Recommends best route based on your trip timeline
  • Analyzes budget adherence and major cost drivers
  • Suggests practical cutback levers when over budget
  • Quantifies impact of daily spending pace
  • Proposes realistic day sequencing to reduce backtracking
  • Flags overpacked days and suggests buffer time
  • Validates feasibility of planned activities
  • Recommends decisions that reduce coordination overhead
  • Suggests explicit owner and deadline for next actions
  • Helps clarify approval workflows

​
Agent planner (trip creation)

When you start a new trip, the Agent planner helps you build an initial draft through conversation.
1

Share trip basics

Tell the agent your destination, travel dates (or month), traveler count, and must-do experiences.Example: “Planning a solo trip to Tokyo in April, budget around $3000, want to visit museums and try authentic ramen.”
2

Review draft plan

The agent generates a structured draft including:
  • Destination and country
  • Start and end dates (inferred from conversation)
  • Traveler count
  • Budget estimate
  • Activities list
  • Day-by-day itinerary skeleton
The draft appears in real-time as PlannerDraft in the response.
3

Refine and apply

Ask follow-up questions to adjust pacing, budget, or activities. Once satisfied, apply the draft to create your trip.

​
What the agent infers

The planner extracts structured data from your conversation:
  • Destination: Inferred from phrases like “to Paris” or “in Barcelona”
  • Dates: Detects YYYY-MM-DD patterns or month references
  • Travelers: Recognizes “solo”, “2 people”, “family of 4”
  • Budget: Parses amounts like “budget $2000” or “spend 1500”
  • Activities: Pulls from must-do experiences you mention

​
AI system prompts

TripLoom uses model selection to choose the appropriate LLM based on prompt complexity and conversation length. The system prompt is built dynamically for each page.

​
Core principles

From backend/internal/ai/prompt.go:10-31:
  • Read-only assistant: Never claims to modify bookings, itinerary, or finance data
  • Transparent about uncertainty: States when data is missing, stale, or degraded
  • Page-aware guidance: Uses pageKey to provide contextual recommendations
  • Concise and decision-oriented: Focuses on tradeoffs and next steps
  • Natural tone: Avoids robotic templates, matches user energy

​
Degraded mode

When live data sources fail (e.g., flight status API is down), the AI enters degraded mode. It continues to provide guidance but prepends a confidence note and avoids overconfident language. Example fallback from backend/internal/ai/service.go:258-278: 
For Berlin to Munich, I can't fetch live transit suggestions right now, but FlixBus is commonly an option on this corridor. Next actions: 1) check FlixBus for your exact date/time, 2) compare against rail duration/price, 3) pick the best reliability-cost tradeoff.

​
Real-time context refresh

The AI can fetch live data to enhance recommendations. This happens automatically when you enable refresh mode in your chat request.

​
Supported real-time sources

From backend/internal/ai/service.go:322-369:
Page KeySourceRequired InputExample
flightsnext_flight_statusFlight number + date”AA1234 2025-06-15”
transitnext_transit_suggestOrigin + destination”from Berlin to Munich”
financefinance_guardrailCurrent trip totalsAuto-computed
itineraryitinerary_riskCurrent snapshotAuto-derived
If required inputs are missing, the AI skips the real-time fetch and provides guidance based on trip context alone.

​
Suggested actions

Every AI response includes suggested actions tailored to the current page. These are hardcoded recommendations from backend/internal/ai/service.go:394-405. Examples:
  • Flights: “Share exact flight number and date for live status”, “Compare price-time tradeoff before selecting”
  • Transit: “Provide ‘from X to Y’ for route options”, “Save top route to itinerary notes”
  • Finance: “Review highest-spend category”, “Set per-day budget target”

​
Using the AI panel

The AI panel is accessible from every trip page in frontend/components/trips/trip-ai-panel.tsx.
1

Open the panel

Click the AI icon in the bottom-right corner of any trip page. The panel slides in from the right.
2

Ask a question

Type your question or request in the text area. Use Cmd+Enter (Mac) or Ctrl+Enter (Windows) to send.Example prompts:
  • “What is my next best step?”
  • “Compare FlixBus vs rail from Berlin to Munich”
  • “Am I over budget?”
3

Review context used

After the AI responds, expand the Context used section to see which data sources were consulted:
  • trip_db_context: Core trip data (destination, dates, budget)
  • next_flight_status: Live flight status (if applicable)
  • next_transit_suggest: Live transit options (if applicable)
Each source shows its status (ok, error, skipped_missing_inputs) and fetch timestamp.

​
Context passed to AI

From frontend/components/trips/trip-ai-panel.tsx:40-46, the AI receives:
  • tripId: Current trip identifier
  • pageKey: Current page (e.g., “flights”, “finance”)
  • pageContext: Page-specific metadata (title, subtitle, path, details)
  • messages: Full conversation history
  • refresh: Boolean to enable real-time data fetching

​
Best practices

  • Be specific: Include flight numbers, dates, and route details for best results
  • Ask for tradeoffs: Instead of “best flight”, try “compare price vs duration for morning flights”
  • Check sources: Review the Context used section to understand data freshness
  • Use natural language: The AI is trained to handle conversational input, not rigid commands
  • Iterate: Follow up with clarifying questions to refine recommendations

​
Limitations

  • Read-only: The AI cannot modify your trip data, bookings, or expenses
  • No booking confirmations: It will never claim to have made a reservation or purchase
  • Live data availability: Real-time sources may be unavailable (degraded mode activates automatically)
  • Token limits: Very long conversations may hit model context limits

​
Technical implementation

The AI service is implemented in Go at backend/internal/ai/service.go and uses:
  • Model selection: ModelSelector chooses the LLM based on prompt characteristics
  • Conversation storage: All messages are persisted in the database via AIRepository
  • Audit logging: Every AI interaction is logged with tripId, pageKey, model, and degraded status
  • Context snapshots: Page context is saved to enable future conversation replay
Frontend integration at frontend/components/trips/trip-ai-panel.tsx handles:
  • Real-time streaming (if enabled)
  • Message display with role-based styling
  • Error handling and retry logic
  • Source attribution and timestamp formatting

Build docs developers (and LLMs) love

Get started for freeTalk to us