How MOSTRAmi Uses Artificial Intelligence
AI-Generated Content
Explanations, chat replies, and voice narrations are generated or assisted by Clio. The current stack uses Gemma 4 E2B/E4B through LiteRT-LM 0.16 for compatible on-device profiles, Gemini 3.6 for the general grounded cloud path, and a dedicated Gemini Live native-audio model when the visitor chooses Live cloud mode. The app provides AI disclosure from onboarding and is designed to support applicable EU AI Act transparency duties and Italian Law 132/2025; this is not a legal certification.
Source Visibility and Grounding
The v0.8 cloud path uses a structured generator, deterministic claim-to-source citation validation and a separate semantic verifier. The chat can display source chips, a grounding label and task type for each assistant message. These controls make provenance inspectable and reduce unsupported output, but they do not guarantee correctness.
The current repository includes one signed, independent Botticelli reference pack backed by linked public sources and an explicit non-endorsement disclosure. Institution-controlled packages, production publishing and museum retrieval remain partner and release gates.
Interactive Artwork Details
The chat may use the captured artwork frame to support pinch-to-zoom, tap-to-focus, detected points of interest, or circle-selected detail crops. These interactions are used to focus the next AI answer on a chosen visible area; they do not automatically publish or retain the photo unless the visitor has separately opted into saving images.
Clio Live Audio
Clio Live Cloud uses Firebase AI Logic and Gemini Live native audio for continuous listening, spoken responses, and interruption support. In the current build, Clio Live can appear as a floating assistant across the main app surfaces. TTS actions are hidden while Live is active, and Live audio stops traditional speech synthesis to avoid overlapping voices. If the Live service is unavailable, the app shows the runtime error and offers retry. When the offline package is installed, the app can fall back to local voice flows.
External Channels and TourSync Direction
TourSync code exists in the Android project, while external messaging access remains a product concept. WhatsApp/Telegram operation, licensing and production delivery are not part of v0.8.0:
- AI identifiability requirement: any future messaging channel must label Clio as an AI system, preserve source visibility and pass privacy, platform and legal review before deployment.
- TourSync testing: group synchronization and profile steering require controlled device and venue validation. The current documentation does not claim a production deployment or that local processing removes every behavioral risk.
Possible Errors
AI models can produce incomplete or inaccurate interpretations. MOSTRAmi is an assisted cultural guide, not an official curatorial source.
Reporting Content
Every AI response in the chat can be reported by the user directly in the app. Reporting categories include offensive or harmful content, misleading inaccuracies, and privacy or copyright concerns.
Cloud and On-Device Processing
Where available, users can select an on-device privacy mode. Cloud functions route through the MOSTRAmi proxy secured by App Check, abuse controls, and cost limits. If cloud capacity is temporarily unavailable, MOSTRAmi attempts local fallback when the offline model is installed; otherwise it asks the user to download offline assets or retry cloud later.