AgentSkillsCN

adk-streaming

精通ADK在Python中实现双向（Bidi）流式传输，以及与Gemini Live API的集成。可用于打造低延迟的语音/视频交互，以及多模态AI智能体的应用。

SKILL.md
--- frontmatter
name: adk-streaming
description: Expert in ADK bidirectional (Bidi) streaming and Gemini Live API integration in Python. Use for implementing low-latency voice/video interaction and multimodal AI agents.

ADK Streaming Specialist

Philosophy & Architecture

Bidi-streaming (Live mode) adds low-latency voice and video interaction. Users can interrupt the agent, and agents can process text, audio, and video inputs in real-time.

Key Components

  • •LiveRequestQueue: Upstream flow for text, audio, and video.
  • •run_live(): Processing events, transcriptions, and multi-agent workflows.
  • •RunConfig: Configure response modalities and context compression.

Implementation Workflow

  1. •Setup: Use a FastAPI-based server for WebSocket communication.
  2. •Input: Stream multimodal data via LiveRequestQueue.
  3. •Handling: Process events from run_live() for real-time reactions.
  4. •Tools: Use "Streaming Tools" for agents to react to intermediate results (e.g., video changes).

Best Practices

  • •Use gemini-2.0-flash for low-latency live interactions.
  • •Implement Voice Activity Detection (VAD) for natural turn-taking.
  • •Read references/streaming.md for part-by-dev-guide series.

Success Criteria

  • •Valid implementation of real-time event loops.
  • •Successful handling of audio/video buffers.
  • •Low-latency response generation with interruption support.