AI Training — How Your AI Learns Everything

How AI Training Works

Vanio AI learns about your properties through 4 manual training methods and 11+ automated learning systems that run in the background. The more the AI knows, the more guest questions it handles autonomously.

Open the training modal from any listing page by clicking Train AI, or go to Settings > AI to see training scores across all listings.

---

The Training Modal

Every listing has a training modal (?train=<listingId>) with three sections:

• Left sidebar — Training progress by category, article counts, and an overall training score
• Center — Smart guidance banner + 4 training method cards + all KB articles organized by category
• Right sidebar — VanioChat with two modes: "Training Assistant" (helps draft articles) and "Guest Demo" (simulates guest conversations to test your AI)

The system recommends what to do next based on what you have already completed.

---

4 Manual Training Methods

Step 1: Mine from Conversations

If you have existing guest conversations (5+), this is the fastest way to train. The AI reads your message history, extracts reusable factual information, and generates KB articles.

1. Click Mine from Conversations in the training modal
2. The system analyzes all conversations for this listing
3. Review the draft articles — each shows the extracted knowledge
4. Select the ones you want to keep and save

The system automatically deduplicates against existing articles (using embedding similarity) and consolidates group-level knowledge if you have multiple units in the same building.

Step 2: Videos & Photos

Upload a walkthrough video of your property or select from existing photos. The AI transcribes the video, identifies knowledge topics, and generates articles.

Video upload:

1. Click Videos & Photos in the training modal
2. Upload an MP4, MOV, or WebM file (up to 500MB)
3. The AI transcribes the video and extracts topics
4. Review the generated articles — toggle "embed video in article" if you want guests to see the video
5. Save

Photo analysis:

1. Switch to the Photos tab
2. Select property photos
3. The AI analyzes what it sees and generates relevant articles (room descriptions, amenities, layout)
4. Review and save

Step 3: Quick Questionnaire

A guided form with ~60 property-specific questions across 10 categories. Answer the questions; the AI generates KB articles from your answers.

Categories:

• Check-in/Check-out (access method, lockbox, directions, checkout instructions)
• Amenities (WiFi, kitchen, laundry, climate, entertainment, pool, gym, EV charging, workspace, and more)
• House Rules (garbage, visitors, other rules)
• Property Information (floor, bedrooms, bathrooms, parking, special features)
• Travel Planning (transport, grocery, restaurants, attractions, airport, medical)
• Technical Issues (WiFi troubleshooting, heating, hot water, appliance quirks, lockout, emergency)
• Customer Service (host contact, building manager, maintenance, local services)
• Maintenance/Housekeeping (cleaning schedule, mid-stay clean, linen, supplies)
• Booking/Reservation (cancellation, minimum stay, payment, modification, deposit, extra guest fees)
• Procedures (early check-in, late checkout, refunds, complaints, escalation, upsells, maintenance)
• Guest Portal (portal link, early/late checkout limits, mid-stay cleaning, invoices) — only shown when the listing has an active portal

Questions are filtered by your listing's amenities — if you don't have a pool, pool questions don't appear. Some questions are pre-filled from structured data (Airbnb WiFi, directions, etc.).

Step 4: Upload Documents

Upload existing house guides, SOPs, or rules documents. Supported formats: PDF, TXT, DOCX, Markdown.

1. Click Upload Documents
2. Drag and drop or browse for files
3. The AI extracts text, splits it into discrete topics, and creates KB articles
4. Review and save

---

11 Automated Learning Systems

These run in the background without any action from you. The AI gets smarter passively as you operate.

Shadow Learning

When the AI escalates a message to you (or answers with low confidence) and you reply to the guest, the system captures your response, pairs it with the guest's question, and creates a KB article. Your real answers become the AI's future responses.

Ratings Feedback (Thumbs Up / Thumbs Down)

Every AI response has a thumbs-up/thumbs-down button in the timeline. Thumbs-up validates the response and reinforces confidence. Thumbs-down flags it for correction and feeds the learning loop.

Supervisor Learning

The AI has a quality supervisor that reviews every response. When the supervisor repeatedly modifies or retries responses for the same reason, the system detects the pattern and auto-creates a KB article to prevent the mistake. Runs nightly.

Rephrase Detection

When a guest rephrases the same question within 30 minutes, the system marks the first AI response as "potentially incomplete." This feeds shadow learning — even if you don't manually flag it, the system knows the AI's answer wasn't good enough.

Repeated Questions Detection

Weekly scan that clusters recent guest questions by topic. If 3+ guests ask the same unanswered question, the system creates a training suggestion: "Multiple guests are asking about X — please provide an answer."

Gap Filler

When the accuracy evaluation finds categories where the AI scores poorly, Gap Filler searches your conversation history for host answers to those topics. If found, it auto-creates articles. If not found, it creates training suggestions asking you to fill the gap.

Implicit Validation

When an AI response gets no negative signal within 24 hours (no thumbs-down, no guest rephrase, and the conversation continued), the system marks it as implicitly validated. This feeds confidence calibration at scale.

Confidence Calibration

Tracks AI stated confidence vs actual outcomes. After enough data points per listing, the system computes calibration curves and recommends dynamic escalation thresholds — so the AI knows more precisely when to handle a question itself vs. when to escalate to you.

Cross-Customer Intelligence

Same building: When a new customer onboards units at an address where another customer already has trained units, the system auto-copies relevant KB articles (shared building info: gym, parking, trash, etc.).

Same region: For new listings in a city with existing customers, the system creates training suggestions based on common regional question patterns (local transport, popular restaurants, etc.).

Procedure Pattern Detection

When the AI handles a situation the same way 5+ times in 30 days (same sequence of tool calls), it detects the pattern and suggests saving it as an AI Skill: "I notice I handle early check-in requests the same way every time — want to save this as a procedure?"

Memory Service (Host Brain)

Three-layer memory that learns how you work:

• Host preferences — stable facts about your style and decisions
• Conversation summaries — compressed digests of each dashboard conversation
• Behavioral patterns — aggregated patterns from your actions (updated nightly)

The AI uses this to adapt its tone, decision-making, and recommendations to match how you operate.

---

Training Dashboard

Go to Settings > AI > Training for the global view:

• Stats — Total learning events, KB articles learned, validated responses, supervisor corrections (7d / 30d / 90d)
• Training Suggestions — Questions the AI needs help with. Type an answer and click "Teach" to create a KB article instantly.
• Learning Events Timeline — Chronological feed of everything the AI learned: shadow learning, ratings, supervisor corrections, auto-created articles, accuracy changes
• Knowledge Base Growth — Visual breakdown of article sources: manually created, AI-generated, from chat history, imported
• Accuracy Trend Chart — Line chart showing accuracy score (0-3) over time per listing
• Per-Listing Scorecards — Accuracy and coverage scores per listing, with question counts

Training Score

Every listing has a training score based on questionnaire coverage: (covered applicable questions / total applicable questions) × 100. View scores on the Settings > AI page to see which listings need more training.

Testing Your AI

Two ways to test:

1. Guest Demo (training modal right sidebar) — Chat with the AI as a guest would, for the specific listing you're training
2. Test AI Lab (/test-ai) — Select any listing + reservation, then chat with the AI to test responses against your KB

Tips

• Start with Mine from Conversations if you have history — it's the fastest path to a well-trained AI
• Upload a walkthrough video — 5 minutes of narration generates dozens of articles
• Use the questionnaire to fill remaining gaps — the coverage score tells you what's missing
• Check Training Suggestions weekly — answering 5 questions takes 2 minutes and significantly improves accuracy
• Watch the accuracy trend — a rising line means your training is working
• The AI learns passively from your normal operations — shadow learning, implicit validation, and supervisor learning all happen without any effort from you