--- title: "LLMHandoff" type: concept tags: ["llm", "agent", "handoff", "transcription", "summarization"] last_updated: 2026-05-02 --- # LLMHandoff(LLM 交接协议) ## Definition LLM Handoff 是在 Agent 管道中,将结构化数据(转录文本、会议记录、文档)传递给下游 LLM 进行摘要/问答/行动项提取的标准接口。定义格式化规则和任务指令,使下游 LLM 能准确理解输入并产生期望输出。 ## Purpose 在 [[StructuredTranscriptJSON]] 基础上,LLM Handoff 定义: 1. **格式化规则**:如何将带时间戳的段落转为 LLM 可读的文本格式 2. **任务指令**:不同任务(摘要/问答/行动项)对应的 prompt 指令 3. **Schema 输出**:LLM 应返回的结构化格式(如 action_items、summary 等) ## Handoff Payload Format ```json { "task": "summarize", "source_type": "transcript", "source_id": "meeting-2026-05-02-001", "total_duration": 3600.5, "speakers": ["SPEAKER_00", "SPEAKER_01"], "content": "[0.0s] Hello, welcome...\n[5.2s] Thank you...\n...", "instructions": { "summarize": "Produce a concise summary, section headers for topic changes, and a bulleted action items list with speaker attribution.", "action_items": "Extract all action items and commitments with the speaker who made them and the timestamp.", "qa": "Answer questions about the transcript using only information present in the content. Cite timestamps." } } ``` ## Downstream Integrations - [[LangChain]]:`ConversationalRetrievalChain` / `LLMChain` - CrewAI:`Agent` 接收 JSON payload 作为 context - 自定义 LLM 管道:直接读取 payload 并注入 system prompt ## Critical Design Considerations - **保留时间戳锚点**:`[5.2s]` 格式使 LLM 能引用具体时刻 - **说话人归属**:`[SPEAKER_00]` 前缀使 LLM 能区分不同发言人的观点 - **分块交付**:超长转录可分批传递给 LLM(避免 token 超限) ## Related Concepts - [[StructuredTranscriptJSON]] — Handoff 的输入数据格式 - [[PIIRedaction]] — Handoff 前必须脱敏(避免 LLM 学习 PII) - [[SpeakerDiarization]] — 说话人标签是 Handoff 文本格式化的核心要素 - [[Human-Handoff]] — LLM → 人类的交接(Agentic 系统中的另一方向) ## Related Sources - [[engineering-voice-ai-integration-engineer]]