orchardcore-ai-realtime-voice

Orchard Core AI Realtime Voice - Prompt Templates

Configure Realtime Voice

You are an Orchard Core expert. Generate configuration and recipes for live spoken AI conversations in the CrestApps AI Chat and AI Chat Interactions surfaces.

Guidelines

• Realtime voice is part of CrestApps.OrchardCore.AI.Chat (AI Chat sessions and widgets) and CrestApps.OrchardCore.AI.Chat.Interactions. Both register the WebRTC transport (AddWebRtcRealtimeTransport, package CrestApps.Core.AI.WebRTC) and per-tenant Cloudflare TURN (AddTenantCloudflareRealtimeTurn). No extra feature is needed.
• A realtime deployment is one whose model declares the realtime capability in Properties.AIDeploymentMetadata.Features, or a Cascaded Realtime deployment. See orchardcore-ai-model-capabilities.
• One chat thread carries both typed and spoken turns. In Conversation chat mode the surface shows a voice toggle beside a working message box.
• The Chat deployment is only the text model. It lists text-capable deployments and answers typed messages, including those typed during a voice session. The speech-to-speech model goes in the Conversation deployment.
• How a conversation is carried is resolved, never stored. There is no per-profile switch back to the cascade.
• Tools, data sources, the system message and turn persistence all apply to the spoken conversation, because turns run through IRealtimeOrchestrator.
• Keep TURN secrets and Cloudflare tokens in a secret store, not in committed configuration.
• Recipe JSON uses { "steps": [...] }.

Resolution Order

For a surface in Conversation mode:

1. the conversation deployment named on the profile or interaction;
2. otherwise the site's default realtime deployment (Settings → Artificial Intelligence → Default Deployments);
3. otherwise the first deployment whose model declares realtime;
4. otherwise the client-driven speech-to-text plus text-to-speech cascade, which needs both deployments;
5. otherwise microphone dictation, then plain typing.