第一次學 Kotlin Koog AI 就上手 Day 32：歷史記錄壓縮：優化對話上下文

val strategy = strategy<String, String>("multi-stage-agent") {

    // 第一階段：收集資訊
    val collectInformation by subgraph<String, String> {
        // 收集用戶需求和相關資訊
    }

    // 壓縮歷史，為下一階段準備乾淨的上下文
    val compressHistory by nodeLLMCompressHistory<String>(
        strategy = HistoryCompressionStrategy.WholeHistory
    )

    // 第二階段：分析和決策
    val analyzeAndDecide by subgraph<String, String> {
        // 基於收集的資訊進行分析和決策
    }

    // 連接各階段
    nodeStart then collectInformation then compressHistory then analyzeAndDecide
}

// 基於訊息數量判斷
private suspend fun AIAgentContextBase.shouldCompress(): Boolean {
    return llm.readSession { prompt.messages.size > 20 }
}

// 基於內容長度判斷
private suspend fun AIAgentContextBase.shouldCompressByLength(): Boolean {
    return llm.readSession {
        prompt.messages.sumOf { it.content.length } > 8000
    }
}

// 在策略中使用條件判斷
edge(executeTool forwardTo compressHistory onCondition { shouldCompress() })

val compressHistory by nodeLLMCompressHistory<String>(
    strategy = HistoryCompressionStrategy.WholeHistory
)

val compressHistory by nodeLLMCompressHistory<String>(
    strategy = HistoryCompressionStrategy.FromLastNMessages(10)
)

val compressHistory by nodeLLMCompressHistory<String>(
    strategy = HistoryCompressionStrategy.Chunked(chunkSize = 10)
)

val compressHistory by nodeLLMCompressHistory<String>(
    strategy = HistoryCompressionStrategy.RetrieveFactsFromHistory(
        Concept(
            keyword = "user_preferences",
            description = "使用者的個人偏好設定，包括語言、主題風格等",
            factType = FactType.MULTIPLE
        ),
        Concept(
            keyword = "issue_resolved",
            description = "客戶的問題是否已經解決？",
            factType = FactType.SINGLE
        )
    )
)

val strategy = strategy<String, String>("agent-with-compression") {
    val processRequest by nodeLLMRequest()
    val executeTool by nodeExecuteTool()
    val sendToolResult by nodeLLMSendToolResult()

    // 在策略圖中加入壓縮節點
    val compressHistory by nodeLLMCompressHistory<ReceivedToolResult>(
        strategy = HistoryCompressionStrategy.FromLastNMessages(10)
    )

    // 定義何時觸發壓縮
    edge(executeTool forwardTo compressHistory onCondition {
        llm.readSession { prompt.messages.size > 15 }
    })
    edge(compressHistory forwardTo sendToolResult)
}

class CustomerServiceAgentWithHistoryCompression {

    // 檢查歷史記錄是否過長（超過 10 條訊息就壓縮）
    private suspend fun AIAgentContextBase.shouldCompressHistory(): Boolean {
        return llm.readSession { prompt.messages.size > 10 }
    }

    private val agent = AIAgent(
        executor = simpleOpenAIExecutor(ApiKeyManager.openAIApiKey!!),
        systemPrompt = """
            你是一個專業的客服助手，負責回答客戶問題。
            請用正體中文回應客戶，保持友善和專業的態度。
        """.trimIndent(),
        llmModel = OpenAIModels.CostOptimized.GPT4_1Mini,
        strategy = createStrategy()
    )

    private fun createStrategy() = strategy<String, String>("customer-service-with-compression") {

        // 定義主要處理節點
        val processRequest by nodeLLMRequest()
        val executeTool by nodeExecuteTool()
        val sendToolResult by nodeLLMSendToolResult()

        // 歷史記錄壓縮節點 - 使用 FromLastNMessages 策略
        val compressHistory by nodeLLMCompressHistory<ReceivedToolResult>(
            strategy = HistoryCompressionStrategy.FromLastNMessages(5)
        )

        // 診斷節點 - 展示壓縮前後的狀態
        val diagnosticNode by node<ReceivedToolResult, ReceivedToolResult>("diagnostic") { toolResult ->
            println("📊 === 歷史記錄壓縮觸發 ===")

            // 顯示壓縮將要發生
            val beforeMessages = llm.readSession { prompt.messages.size }
            println("🔍 壓縮前訊息數量: $beforeMessages 條")
            println("⚡ 即將觸發壓縮：保留最近 5 條訊息，將早期對話摘要化")

            toolResult
        }

        // 壓縮後檢查節點
        val postCompressionCheck by node<ReceivedToolResult, ReceivedToolResult>("post_compression") { toolResult ->
            val afterMessages = llm.readSession { prompt.messages.size }
            println("✅ 壓縮完成！目前訊息數量: $afterMessages 條")

            toolResult
        }

        // 建立執行流程
        edge(nodeStart forwardTo processRequest)

        // 如果是助理回應，直接結束
        edge(processRequest forwardTo nodeFinish onAssistantMessage { true })

        // 如果需要使用工具，執行工具
        edge(processRequest forwardTo executeTool onToolCall { true })

        // 執行工具後檢查是否需要壓縮歷史
        edge(executeTool forwardTo diagnosticNode onCondition { shouldCompressHistory() })
        edge(diagnosticNode forwardTo compressHistory)
        edge(compressHistory forwardTo postCompressionCheck)
        edge(postCompressionCheck forwardTo sendToolResult)

        // 如果不需要壓縮，直接發送工具結果
        edge(executeTool forwardTo sendToolResult onCondition { !shouldCompressHistory() })

        // 處理工具結果後的後續動作
        edge(sendToolResult forwardTo executeTool onToolCall { true })
        edge(sendToolResult forwardTo nodeFinish onAssistantMessage { true })
    }

    suspend fun handleCustomerQuery(query: String): String {
        return agent.run(query)
    }
}

// 一般客服：保留最近對話即可
HistoryCompressionStrategy.FromLastNMessages(15)

// 技術支援：需要保留完整上下文
HistoryCompressionStrategy.WholeHistory

// 銷售對話：提取客戶偏好和需求
RetrieveFactsFromHistory(
    Concept("customer_needs", "客戶需求和偏好", FactType.MULTIPLE),
    Concept("budget_range", "客戶預算範圍", FactType.SINGLE)
)

// 長期專案：分段保留不同階段資訊
HistoryCompressionStrategy.Chunked(20)

val compressHistory by nodeLLMCompressHistory<String>(
    strategy = HistoryCompressionStrategy.FromLastNMessages(10),
    preserveMemory = true  // 保留 AgentMemory feature 的記憶訊息
)