Conversational Agents

You're on a voice call with Meggy and it starts explaining something, but you already know the background — you just want the answer. With most AI assistants, you'd have to wait for the response to finish, then rephrase your question. With Meggy's conversational agents, you can simply say "skip that, just tell me the result" and the AI immediately pivots.

Conversational mode adds real-time interrupt detection to any channel that supports it. It transforms the traditional "wait your turn" AI interaction into something that feels like a genuine conversation.

How Interrupts Work

When conversational mode is enabled, Meggy continuously monitors for incoming user input — even while the AI is generating a response. Every incoming message is evaluated by a two-tier classification system:

Tier 1 — Fast Heuristics

Pattern matching runs in under a millisecond, checking for common interrupt signals:

Type Examples What Happens
Correction "no wait", "actually", "I meant…" AI stops, re-generates with context
New Topic "forget that, tell me about…" AI stops, starts fresh
Filler "uh huh", "hmm", "um" Ignored — AI keeps going
Acknowledgment "thanks", "great", "got it" Ignored — AI keeps going

Tier 2 — LLM Re-Classification

When the fast heuristic isn't confident enough, an optional async pass uses a lightweight model (like Gemini Flash) to semantically classify the interrupt. This catches more nuanced cases while keeping false-positive rates low.

Turn Management

A state machine tracks the conversation lifecycle for each active session:

IdleUser SpeakingAI GeneratingIdle

If the user speaks during the "AI Generating" phase, the turn manager detects this, classifies the intent, and if it's a correction or new topic, transitions to the Interrupted state. The current generation is aborted, and a new response is produced using context stitching.

Context Stitching

When an interrupt happens, the AI has already generated a partial response. Context stitching combines what was generated so far with the user's interruption to produce a coherent re-generation:

Voice Barge-In

For voice and phone channels, interrupting the AI means speaking over its voice output. This creates a challenge: the microphone picks up the AI's own voice from the speakers, which could trigger a false interrupt.

Meggy solves this with an echo gate — a real-time audio filter that runs before the VAD (Voice Activity Detection) pipeline:

For phone calls, the same principle applies using µ-law RMS energy detection directly on the raw telephone audio stream.

Sensitivity Levels

You can tune how aggressively Meggy interprets incoming messages as interrupts:

Level Behavior
1 Very conservative — only explicit corrections trigger interrupts
2 Conservative — corrections and clear topic changes
3 Default — balanced for most conversations
4 Aggressive — shorter phrases may trigger interrupts
5 Very aggressive — almost any input during generation interrupts

Supported Channels

Channel Interrupt Support Method
Voice Echo gate + VAD + speech recognition
Phone µ-law energy detection + VAD + STT
Telegram Text messages during generation
Desktop Streaming UI handles flow
WhatsApp Not supported

Enabling Conversational Mode

  1. Open your agent's Channel Bindings settings
  2. Select a channel that supports interrupts (Voice, Phone, or Telegram)
  3. Toggle Conversational Mode on
  4. Adjust the Interrupt Sensitivity slider to your preference
  5. Start talking — and feel free to interrupt