Real-Time Communication

Real-Time Communication is the foundation of the RTC Agent user experience. Built on Centrifuge WebSocket, the system uses a dual-channel architecture: the Topic channel ensures no messages are lost, and the Live channel ensures low-latency streaming output — reliability and speed, without compromising either.

Dual-Channel Architecture

	Topic Channel	Live Channel
Purpose	State change events	Streaming message intermediate chunks
Persistence	✅ Written to database	❌ Redis PUB/SUB
Offset	✅ Strictly incrementing	❌ None
Offline Recovery	✅ Supported	❌ Not supported
Latency	Lower (requires persistence)	Extremely low (in-memory forwarding)

💡 Design principle: Important messages go through Topic, speed-critical messages go through Live. State changes must be delivered reliably, while losing one or two intermediate chunks of streaming output is harmless.

Event Distribution

Different types of events are routed to different channels:

Event Type	Topic	Live	Description
session.created/updated	✅	❌	Session state changes
turn.created/updated	✅	❌	Turn state changes
message.created	✅	❌	New message created
message.updated (stream complete)	✅	❌	Final state after stream completion
message.updated (stream intermediate chunk)	❌	✅	Intermediate fragments of streaming output
rtc.updated	✅	❌	RTC tool call state changes

Offset Mechanism

The Offset is the core of the Topic channel’s reliability — each event is assigned a strictly incrementing sequence number. Clients detect lost messages by checking sequence continuity:

Feature	Description
Strictly incrementing	Each Topic event’s Offset is greater than the previous one
Gap detection	When the client detects an Offset gap, it automatically fetches missing history
No loss, no duplicates	Guarantees messages are neither lost nor duplicated

Streaming Message Flow

AI model streaming output is the most frequently perceived real-time feature by users. The system achieves both real-time and reliable streaming through Topic + Live collaboration:

Phase	Channel	Action
First chunk	Topic	Create message record, push message.created
Intermediate chunks	Live	Append to Redis buffer, push to Live for real-time display
Last chunk	Topic	Concatenate full content, update database, push final state

💡 Why two steps? The Live channel lets each chunk reach the frontend instantly, giving users a smooth typing effect. Finally, the Topic channel pushes the complete message, ensuring that even if Live lost some chunks, the final state is complete and correct.

Offline Recovery

Network connections can’t always be stable. The system provides different strategies for recovery after disconnection:

Scenario	Behavior	User Perception
🟢 Brief disconnection	Messages from offline period are automatically pushed after reconnection	Nearly imperceptible
🟡 Extended disconnection	Offset gap detected, proactively fetches history to fill in	Sees backfilled messages
🔴 History cleaned up	Epoch changed, local Offset cleared, starts from latest position	Continues from current state

Epoch is the timeline identifier for Offsets. When the server performs large-scale cleanup or rebuilding, the Epoch changes and clients need to reset their Offset to start from the latest position — because the old history no longer exists.

Reconnection Mechanism

Mechanism	Description
Auto-reconnect	Centrifuge SDK built-in reconnection with exponential backoff
Token refresh	Automatically refreshes authentication token when connection is restored
Offset persistence	Offset is stored in IndexedDB and restored after page refresh
Fallback strategy	User must re-login when token refresh fails

📌 Offset persistence to IndexedDB means: even if the user closes the browser tab and reopens it, they can continue receiving messages from where they left off, without missing any updates.

Architecture Overview

Next Steps

• Remote Tool Calling — Learn about the core protocol carried by real-time communication
• Session Management — Learn how real-time events are organized within sessions
• Context Management — Learn how messages are intelligently compressed