Communications

Per-property manuals (AI Q&A grounded)

← Back to Communications

Upload property-specific manuals (PDF / Word / images) — the AI uses them to answer guest questions grounded in operator-curated source material.

Canonical use case: a guest asks "how do I use the air conditioning?" — the AI looks up the air-conditioning section in the manual operator uploaded for that property and replies with the actual instructions, not a generic guess.

Where uploads live

V28 → Properties → pick property → Manuals tab.

Each property has its own set of uploaded files. Cross-property sharing is not supported — each manual is scoped to the property it was uploaded under.

Upload flow

  1. Tap + Add manual → file picker.
  2. Accepts PDF, DOCX, JPG, PNG. Max size 20 MB per file.
  3. Optionally tag the manual (e.g. air-conditioning, wifi, parking, appliances).
  4. The platform extracts text + runs embeddings server-side (gh#542 — per-property embeddings).
  5. Embedding completion shows "Indexed" — the manual is now searchable by the AI.

How the AI uses it

When a guest asks a question:

  1. The AI Suggest pipeline pulls the reservation's property.
  2. Runs a vector similarity search over the property's manual embeddings.
  3. The top-K chunks become context for the reply draft.
  4. The draft references the manual content explicitly — no fabricated instructions.

In practice this means a guest's "how do I use the AC?" gets answered with the actual brand + model + button sequence operator wrote, not a generic LLM guess.

What gets indexed

  • PDFs — text extracted page-by-page, chunked, embedded.
  • DOCX — same, paragraph-by-paragraph.
  • Images — OCR'd first (LLM vision pass), then chunked + embedded. Useful for photos of appliance panels with text.

Updating a manual

To replace a manual: delete the old one + upload the new one. The embeddings re-build on the new file.

Compaction (gh#545)

Periodically the platform consolidates message refinements (operator-edited AI suggestions) back into the property manuals as tenant personality + FAQs. This preserves long-tail learning without inflating every live prompt.

Compaction is operator-triggered: tap Compact refinements → preview which message edits get baked in → confirm.

What this is NOT

This is NOT the FAQ Knowledge Base. The FAQ is shorter, hand-curated Q&A entries. The manuals are full source documents the AI retrieves from.

In practice you'll use both: - FAQ for short, recurring questions ("what's the WiFi password?"). - Manuals for long-form reference ("how do I use the dishwasher?").


Implements: gh#542 (per-property manual upload + embeddings — AI Q&A grounded in operator-uploaded manuals), gh#545 (periodic compaction of message refinements into tenant personality + FAQs). Related: FAQ Knowledge Base, AI Cockpit → Prompts.

Source: the FlatsBratislava operator manual.