模型串聯與備援鏈

AI ↔ AI

優先使用快速模型；若信心不足則升級至較大的模型。

5 個節點 · 5 條連接ml-infra

agentsystem

視覺化

接收查詢event

使用者請求進入串聯處理管線。

↓sequential→ 快速模型（Haiku）

快速模型（Haiku）agent

低成本、低延遲的第一次嘗試。

↓sequential→ 信心度檢查

信心度檢查system

依模型信心評分閾值決定路由方向。

↓conditional→ 回傳回應

↓conditional→ 大型模型（Opus）

大型模型（Opus）agent

應對複雜查詢的高能力備援選項。

↓sequential→ 回傳回應

回傳回應api

將最終答案交付給呼叫端。

在編輯器開啟

uc-model-cascade.osop.yaml

osop_version: "1.0"
id: "model-cascade"
name:"模型串聯與備援鏈"
description:"優先使用快速模型；若信心不足則升級至較大的模型。"

nodes:
  - id: "receive"
    type: "event"
    name: "接收查詢"
    description: "使用者請求進入串聯處理管線。"

  - id: "fast_model"
    type: "agent"
    subtype: "llm"
    name: "快速模型（Haiku）"
    description: "低成本、低延遲的第一次嘗試。"
    timeout_sec: 5

  - id: "check_confidence"
    type: "system"
    name: "信心度檢查"
    description: "依模型信心評分閾值決定路由方向。"

  - id: "large_model"
    type: "agent"
    subtype: "llm"
    name: "大型模型（Opus）"
    description: "應對複雜查詢的高能力備援選項。"
    timeout_sec: 30

  - id: "respond"
    type: "api"
    name: "回傳回應"
    description: "將最終答案交付給呼叫端。"

edges:
  - from: "receive"
    to: "fast_model"
    mode: "sequential"
  - from: "fast_model"
    to: "check_confidence"
    mode: "sequential"
  - from: "check_confidence"
    to: "respond"
    mode: "conditional"
    when: "confidence >= 0.8"
  - from: "check_confidence"
    to: "large_model"
    mode: "conditional"
    when: "confidence < 0.8"
  - from: "large_model"
    to: "respond"
    mode: "sequential"