How does interrupt work in TEN-Agent

Overview

The interrupt mechanism inTEN-Agent consists of two main parts: Interrupt Detection and Interrupt Response. This document details both parts and explains how the interrupt command propagates through the AI agent graph.

Part 1: Interrupt Detection

1. Current Interrupt Detection Implementation

The current (interrupt_detector_python) extension implements a text-based interrupt detection mechanism:

def on_data(self, ten: TenEnv, data: Data) -> None:
    text = data.get_property_string(TEXT_DATA_TEXT_FIELD)
    final = data.get_property_bool(TEXT_DATA_FINAL_FIELD)

    # Trigger interrupt when text is final or reaches threshold length
    if final or len(text) >= 2:
        self.send_flush_cmd(ten)

The interrupt detector triggers in the following cases:

1. When receiving final text (is_final = true)

2. When text length reaches a threshold (≥ 2 characters)

2. Customize Interrupt Detection

To implement your own interrupt detection logic, you can refer to the implementation ofinterrupt_detector_python as an example and customize the interrupt conditions based on your specific needs.

Part 2: Interrupt Response

Chain Processing in AI Agent Graph

In a typical AI agent graph, the interrupt command (flush) follows a chain processing pattern:

Interrupt Detector
       ↓
    LLM/ChatGPT
       ↓
      TTS
       ↓
   agora_rtc

Each extension in the chain follows two key steps when receiving a flush command:

1. Clean up its own resources and internal state

2. Forward the flush command to downstream extensions

This ensures that:

• Extensions are cleaned up in the correct order

• No residual data flows through the system

• Each extension returns to a clean state before the next operation

Conclusion

TEN-Agent's interrupt mechanism uses a chain processing pattern to ensure orderly cleanup of all extensions in the AI agent graph. When an interrupt occurs, each extension first cleans up its own state and then forwards the flush command to downstream extensions, ensuring a clean system state for subsequent operations.