Cloud Monitor：make-instance（執行個體構建）

函數預覽

輸入要求

• 上遊節點輸出的任意列資料。

• 必須包含 by 指定的分組鍵欄位。

• 必須包含列定義中引用的所有源欄位。

輸出資料行

行數變化

M → N（M ≥ N）：多行事件彙總為每組一行，輸出行數等於分組數。

未經處理資料（10 條離散事件記錄）

以下為一個使用者請求 AI Agent "分析最近 7 天的錯誤記錄檔"的完整互動過程。Agent 先後調用兩個工具並返回結論，產生 10 條離散事件：

配置樣本

{
  "id": "assemble",
  "type": "make-instance",
  "parameters": {
    "question": "first(question)",
    "answer": "last(output)",
    "model": "any(model)",
    "max_latency": "max(latency_ms)",
    "total_input": "sum(token_input)",
    "total_output": "sum(token_output)",
    "avg_latency": "avg(latency_ms)",
    "llm_calls": "count(model)",
    "tool_count": "count(tool_name)",
    "err_tools": "count_if(tool_success = 'false')",
    "tools": "array_distinct(tool_name)",
    "events": "array(event_type)",
    "tool_chain": "join(tool_name, ' → ')",
    "tool_detail": "json_pack(tool_name, tool_args, tool_success)",
    "by": "session_id,trace_id"
  }
}

處理後（寬表 · 1 行 x 16 列）

10 行窄表彙總為 1 行寬表。對比上方未經處理資料，每一列由指定函數從 10 行中彙總而來：

10 條離散事件彙總為 1 條完整樣本執行個體，覆蓋選值、計算、組合三類函數。缺失欄位自動輸出 NULL。整個過程純 CPU 運算，無 LLM/GPU 依賴。

樣本 1：最簡用法

按 trace 粒度彙總，每列取組內任意非空值。

{
  "id": "assemble",
  "type": "make-instance",
  "parameters": {
    "question": "any(question)",
    "output": "any(output)",
    "model": "any(model)",
    "by": "session_id,trace_id"
  }
}

樣本 2：選值 + 統計 + 打包

{
  "id": "assemble",
  "type": "make-instance",
  "parameters": {
    "question": "first(question)",
    "answer": "last(output)",
    "model": "any(model)",
    "max_latency": "max(latency_ms)",
    "total_tokens": "sum(token_input)",
    "tool_count": "count(tool_name)",
    "tools": "array_distinct(tool_name)",
    "tool_chain": "join(tool_name, ' → ')",
    "by": "session_id,trace_id"
  }
}

樣本 3：完整 Pipeline（執行個體構建 → 清洗 → 採樣 → AI 評估）

過濾有效事件類型 → 資料群組裝 → AI 評估品質。

{
  "nodes": [
    {
      "id": "filter_events",
      "type": "where",
      "parameters": {
        "filter": "event_type IN ('user_query','system_prompt','tool_call','tool_result','assistant_content','completion')"
      }
    },
    {
      "id": "extract", "type": "extend",
      "parameters": {
        "session_id": "json_extract_scalar(attributes, '$.gen_ai.session.id')",
        "span_kind": "json_extract_scalar(attributes, '$.gen_ai.span.kind')",
        "question": "json_extract_scalar(attributes, '$.input.value')",
        "answer": "json_extract_scalar(attributes, '$.output.value')",
        "model": "json_extract_scalar(attributes, '$.gen_ai.request.model')",
        "tool_name": "json_extract_scalar(attributes, '$.gen_ai.tool.name')",
        "input_tokens": "json_extract_scalar(attributes, '$.gen_ai.usage.input_tokens')"
      }
    },
    {
      "id": "filter_events", "type": "where",
      "parameters": { "filter": "span_kind IN ('AGENT','LLM','TOOL')" }
    },
    {
      "id": "assemble", "type": "make-instance",
      "parameters": {
        "question": "first(question)",
        "answer": "last(answer)",
        "model": "last(model)",
        "total_tokens": "sum(input_tokens)",
        "tools": "array_distinct(tool_name)",
        "tool_chain": "join(tool_name, ' → ')",
        "by": "session_id,traceId"
      }
    },
    { "id": "filter_empty", "type": "where", "parameters": { "filter": "question IS NOT NULL AND length(question) > 0" } },
    { "id": "exact", "type": "dedup-exact", "parameters": { "field": "question" } },
    { "id": "fuzzy", "type": "dedup-fuzzy", "parameters": { "field": "question", "threshold": "3" } },
    { "id": "take", "type": "sample", "parameters": { "n": 50 } },
    { "id": "eval", "type": "llm-call", "parameters": { "prompt": "@eval/quality.md", "fields": "question,answer", "format": "json", "as": "eval" } },
    { "id": "stats", "type": "doc-stats", "parameters": { "field": "question" } }
  ]
}

樣本 4：OT Trace 實戰 — 兩步轉換 + 全文整合

實戰情境：28 條 OT Span（AGENT/LLM/TOOL/EXTERNAL 等）→ 過濾無關事件 → 組裝為 1 行寬表 → 合并為 full_text 列供下遊 AI 評估。

{
  "nodes": [
    {
      "id": "extract", "type": "extend",
      "parameters": {
        "session_id": "json_extract_scalar(attributes, '$.gen_ai.session.id')",
        "span_kind": "json_extract_scalar(attributes, '$.gen_ai.span.kind')",
        "model": "json_extract_scalar(attributes, '$.gen_ai.request.model')",
        "input_tokens": "json_extract_scalar(attributes, '$.gen_ai.usage.input_tokens')",
        "output_tokens": "json_extract_scalar(attributes, '$.gen_ai.usage.output_tokens')",
        "tool_name": "json_extract_scalar(attributes, '$.gen_ai.tool.name')",
        "tool_args": "json_extract_scalar(attributes, '$.gen_ai.tool.call.arguments')",
        "input_value": "json_extract_scalar(attributes, '$.input.value')",
        "output_value": "json_extract_scalar(attributes, '$.output.value')",
        "agent_id": "json_extract_scalar(attributes, '$.agent.id')",
        "dur_ms": "cast(duration as bigint) / 1000000"
      }
    },
    {
      "id": "filter_events", "type": "where",
      "parameters": { "filter": "span_kind IN ('AGENT','LLM','TOOL')" }
    },
    {
      "id": "assemble", "type": "make-instance",
      "parameters": {
        "question": "min_by(input_value, startTime)",
        "answer": "max_by(output_value, endTime)",
        "model": "max_by(model, endTime)",
        "total_input_tokens": "sum(cast(input_tokens as bigint))",
        "total_output_tokens": "sum(cast(output_tokens as bigint))",
        "llm_calls": "count_if(span_kind = 'LLM')",
        "tool_calls": "count_if(span_kind = 'TOOL')",
        "e2e_latency": "max(dur_ms)",
        "models": "array_distinct(model)",
        "tools": "array_distinct(tool_name)",
        "tool_chain": "join(tool_name, ' → ')",
        "by": "session_id,traceId"
      }
    },
    {
      "id": "compose", "type": "extend",
      "parameters": {
        "full_text": "concat('## 使用者輸入', chr(10), substr(question, 1, 500), chr(10), chr(10), '## 最終輸出', chr(10), answer)"
      }
    },
    { "id": "filter_empty", "type": "where", "parameters": { "filter": "question IS NOT NULL AND length(question) > 0" } },
    { "id": "eval", "type": "llm-call", "parameters": { "prompt": "@eval/quality.md", "fields": "full_text", "format": "json", "as": "eval" } }
  ]
}

第一步 extend + where + make-instance 完成欄位提取、事件過濾和資料群組裝（28 行 → 1 行）；第二步 extend 將 question + answer 合并為 full_text 供 llm-call 做整體評估。全過程純 CPU。

Pipeline 整合建議

• 將 make-instance 放在 Pipeline 最前端（或 extend 欄位提取之後），作為事件級到樣本級的轉換入口。

• 後接 where 過濾空值行，再進入 dedup / sample / llm-call。

• 對 OT Trace 資料，先用 extend 從 attributes JSON 中提取扁平欄位，再用 make-instance 彙總。

• 選值類函數足以覆蓋 80% 情境；統計類和組合類別按需添加。

• 高階使用者可直接寫 SQL 彙總運算式，與文法糖函數混用。

邊界行為