From 9902ca3563133d434f27eaa77e864c5c06b5b332 Mon Sep 17 00:00:00 2001
From: Losita <2810873701@qq.com>
Date: Sun, 3 May 2026 23:21:03 +0800
Subject: [PATCH] Version: 0.9.65.dev.260503
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

后端：
1. 阶段 1.5/1.6
收口 llm-service / rag-service，统一模型出口与检索基础设施入口，清退 backend/infra/llm 与 backend/infra/rag 旧实现；
2. 同步更新相关调用链与微服务迁移计划文档
---
 .../feedbacklocate/service.go                 |  16 +--
 backend/active_scheduler/selection/service.go |  16 +--
 backend/cmd/start.go                          |  50 ++++----
 backend/memory/module.go                      |  18 +--
 .../orchestrator/llm_decision_orchestrator.go |  18 +--
 .../orchestrator/llm_write_orchestrator.go    |  14 +--
 backend/memory/service/read_scope.go          |   6 +-
 backend/memory/service/read_service.go        |   8 +-
 backend/memory/vectorsync/syncer.go           |  12 +-
 backend/memory/worker/decision_flow.go        |   6 +-
 backend/memory/worker/runner.go               |   6 +-
 backend/newAgent/model/graph_run_state.go     |  18 +--
 backend/newAgent/node/chat.go                 |  32 ++---
 backend/newAgent/node/deliver.go              |   8 +-
 .../newAgent/node/execute/action_router.go    |   6 +-
 backend/newAgent/node/execute/run.go          |   4 +-
 backend/newAgent/node/plan.go                 |  14 +--
 backend/newAgent/node/quick_task.go           |   8 +-
 backend/newAgent/node/unified_compact.go      |   4 +-
 backend/newAgent/prompt/compact_msg1.go       |   6 +-
 backend/newAgent/prompt/compact_msg2.go       |   6 +-
 backend/newAgent/router/decision_parser.go    |   4 +-
 backend/newAgent/shared/node_thinking.go      |   8 +-
 .../newAgent/shared/node_unified_compact.go   |   4 +-
 backend/newAgent/stream/emitter.go            |   6 +-
 backend/newAgent/tools/registry.go            |   4 +-
 backend/service/agent_bridge.go               |  10 +-
 backend/service/agentsvc/agent.go             |  25 ++--
 backend/service/agentsvc/agent_meta.go        |  34 +++---
 backend/service/agentsvc/agent_newagent.go    |  23 ++--
 .../service/agentsvc/agent_stream_fallback.go |  28 +++--
 backend/service/agentsvc/reasoning_summary.go |  12 +-
 backend/service/course.go                     |   6 +-
 backend/service/course_parse_ark.go           |  20 ++--
 backend/{infra => services}/llm/ark.go        |  18 +--
 .../{infra => services}/llm/ark_adapter.go    |  26 ++--
 .../llm/ark_responses_client.go               |  21 ++--
 backend/{infra => services}/llm/client.go     |  79 +++----------
 backend/{infra => services}/llm/json.go       |  20 +---
 backend/services/llm/service.go               | 109 +++++++++++++++++
 .../rag/service.go => services/rag/api.go}    |   2 +-
 .../rag/chunk/text_chunker.go                 |   2 +-
 .../{infra => services}/rag/config/config.go  |   0
 .../{infra => services}/rag/core/errors.go    |   0
 .../rag/core/interfaces.go                    |   0
 .../{infra => services}/rag/core/observer.go  |   4 +-
 .../{infra => services}/rag/core/pipeline.go  |   0
 backend/{infra => services}/rag/core/types.go |   0
 .../{infra => services}/rag/corpus/common.go  |   0
 .../rag/corpus/memory_corpus.go               |   2 +-
 .../rag/corpus/web_corpus.go                  |   2 +-
 .../rag/embed/eino_embedder.go                |   0
 .../rag/embed/mock_embedder.go                |   0
 backend/{infra => services}/rag/factory.go    |  12 +-
 backend/{infra => services}/rag/observe.go    |   2 +-
 backend/{infra => services}/rag/rag.go        |  10 +-
 .../rag/rerank/eino_reranker.go               |   2 +-
 .../rag/rerank/noop_reranker.go               |   2 +-
 .../rag/retrieve/vector_retriever.go          |   2 +-
 backend/{infra => services}/rag/runtime.go    |   8 +-
 backend/services/rag/service.go               | 111 ++++++++++++++++++
 .../rag/store/inmemory_store.go               |   2 +-
 .../rag/store/milvus_store.go                 |   2 +-
 .../rag/store/vector_store.go                 |   2 +-
 .../微服务四步迁移与第二阶段并行开发计划.md   |  26 +++-
 65 files changed, 550 insertions(+), 376 deletions(-)
 rename backend/{infra => services}/llm/ark.go (73%)
 rename backend/{infra => services}/llm/ark_adapter.go (68%)
 rename backend/{infra => services}/llm/ark_responses_client.go (93%)
 rename backend/{infra => services}/llm/client.go (56%)
 rename backend/{infra => services}/llm/json.go (67%)
 create mode 100644 backend/services/llm/service.go
 rename backend/{infra/rag/service.go => services/rag/api.go} (97%)
 rename backend/{infra => services}/rag/chunk/text_chunker.go (96%)
 rename backend/{infra => services}/rag/config/config.go (100%)
 rename backend/{infra => services}/rag/core/errors.go (100%)
 rename backend/{infra => services}/rag/core/interfaces.go (100%)
 rename backend/{infra => services}/rag/core/observer.go (97%)
 rename backend/{infra => services}/rag/core/pipeline.go (100%)
 rename backend/{infra => services}/rag/core/types.go (100%)
 rename backend/{infra => services}/rag/corpus/common.go (100%)
 rename backend/{infra => services}/rag/corpus/memory_corpus.go (98%)
 rename backend/{infra => services}/rag/corpus/web_corpus.go (98%)
 rename backend/{infra => services}/rag/embed/eino_embedder.go (100%)
 rename backend/{infra => services}/rag/embed/mock_embedder.go (100%)
 rename backend/{infra => services}/rag/factory.go (91%)
 rename backend/{infra => services}/rag/observe.go (92%)
 rename backend/{infra => services}/rag/rag.go (60%)
 rename backend/{infra => services}/rag/rerank/eino_reranker.go (86%)
 rename backend/{infra => services}/rag/rerank/noop_reranker.go (91%)
 rename backend/{infra => services}/rag/retrieve/vector_retriever.go (97%)
 rename backend/{infra => services}/rag/runtime.go (97%)
 create mode 100644 backend/services/rag/service.go
 rename backend/{infra => services}/rag/store/inmemory_store.go (98%)
 rename backend/{infra => services}/rag/store/milvus_store.go (99%)
 rename backend/{infra => services}/rag/store/vector_store.go (75%)

diff --git a/backend/active_scheduler/feedbacklocate/service.go b/backend/active_scheduler/feedbacklocate/service.go
index 6fa5897..4e07fcf 100644
--- a/backend/active_scheduler/feedbacklocate/service.go
+++ b/backend/active_scheduler/feedbacklocate/service.go
@@ -11,7 +11,7 @@ import (
 
 	"github.com/LoveLosita/smartflow/backend/active_scheduler/ports"
 	"github.com/LoveLosita/smartflow/backend/active_scheduler/trigger"
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 )
 
 const locateMaxTokens = 800
@@ -24,7 +24,7 @@ const locateMaxTokens = 800
 // 3. 不创建新工具系统，也不直接产出 preview。
 type Service struct {
 	reader ports.ScheduleReader
-	client *infrallm.Client
+	client *llmservice.Client
 	clock  func() time.Time
 	logger *log.Logger
 }
@@ -34,7 +34,7 @@ type Service struct {
 // 说明：
 // 1. reader / client 允许为空，方便在模型不可用或读模型暂时不可用时直接回退 ask_user。
 // 2. 真正的定位能力只在 Resolve 内部按需启用。
-func NewService(reader ports.ScheduleReader, client *infrallm.Client) *Service {
+func NewService(reader ports.ScheduleReader, client *llmservice.Client) *Service {
 	return &Service{
 		reader: reader,
 		client: client,
@@ -101,15 +101,15 @@ func (s *Service) Resolve(ctx context.Context, req Request) (Result, error) {
 		return s.buildAskUserResult(req, "定位 prompt 构造失败"), nil
 	}
 
-	messages := infrallm.BuildSystemUserMessages(strings.TrimSpace(locateSystemPrompt), nil, userPrompt)
-	resp, rawResult, err := infrallm.GenerateJSON[llmResponse](
+	messages := llmservice.BuildSystemUserMessages(strings.TrimSpace(locateSystemPrompt), nil, userPrompt)
+	resp, rawResult, err := llmservice.GenerateJSON[llmResponse](
 		ctx,
 		s.client,
 		messages,
-		infrallm.GenerateOptions{
+		llmservice.GenerateOptions{
 			Temperature: 0.1,
 			MaxTokens:   locateMaxTokens,
-			Thinking:    infrallm.ThinkingModeDisabled,
+			Thinking:    llmservice.ThinkingModeDisabled,
 			Metadata: map[string]any{
 				"stage":           "active_scheduler_feedback_locate",
 				"candidate_count": len(candidates),
@@ -340,7 +340,7 @@ func cloneAndTrimStrings(values []string) []string {
 	return result
 }
 
-func truncateRaw(raw *infrallm.TextResult) string {
+func truncateRaw(raw *llmservice.TextResult) string {
 	if raw == nil {
 		return ""
 	}
diff --git a/backend/active_scheduler/selection/service.go b/backend/active_scheduler/selection/service.go
index 4f9bf11..9f52d52 100644
--- a/backend/active_scheduler/selection/service.go
+++ b/backend/active_scheduler/selection/service.go
@@ -10,7 +10,7 @@ import (
 	"time"
 
 	"github.com/LoveLosita/smartflow/backend/active_scheduler/candidate"
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 )
 
 const selectionMaxTokens = 1200
@@ -22,7 +22,7 @@ const selectionMaxTokens = 1200
 // 2. LLM 失败、输出非法或选择不存在候选时，回退到后端 fallback candidate；
 // 3. 不写 preview、不发通知、不修改正式日程。
 type Service struct {
-	client *infrallm.Client
+	client *llmservice.Client
 	clock  func() time.Time
 	logger *log.Logger
 }
@@ -33,7 +33,7 @@ type Service struct {
 // 1. client 允许为空；为空时选择器只走确定性 fallback，便于本地测试和降级；
 // 2. 真正的模型接入在 cmd/start.go 中完成：aiHub.Pro -> llm.Client -> selection.Service；
 // 3. 选择器本身不持有模型配置，只表达本业务域的 prompt 和结果校验。
-func NewService(client *infrallm.Client) *Service {
+func NewService(client *llmservice.Client) *Service {
 	return &Service{
 		client: client,
 		clock:  time.Now,
@@ -70,19 +70,19 @@ func (s *Service) Select(ctx context.Context, req SelectRequest) (Result, error)
 		return buildFallbackResult(req, "选择器 prompt 构造失败: "+err.Error()), nil
 	}
 
-	messages := infrallm.BuildSystemUserMessages(
+	messages := llmservice.BuildSystemUserMessages(
 		strings.TrimSpace(selectionSystemPrompt),
 		nil,
 		userPrompt,
 	)
-	resp, rawResult, err := infrallm.GenerateJSON[llmSelectionResponse](
+	resp, rawResult, err := llmservice.GenerateJSON[llmSelectionResponse](
 		ctx,
 		s.client,
 		messages,
-		infrallm.GenerateOptions{
+		llmservice.GenerateOptions{
 			Temperature: 0.1,
 			MaxTokens:   selectionMaxTokens,
-			Thinking:    infrallm.ThinkingModeDisabled,
+			Thinking:    llmservice.ThinkingModeDisabled,
 			Metadata: map[string]any{
 				"stage":           "active_scheduler_select",
 				"candidate_count": len(req.Candidates),
@@ -275,7 +275,7 @@ func firstNonEmpty(values ...string) string {
 	return ""
 }
 
-func truncateRaw(raw *infrallm.TextResult) string {
+func truncateRaw(raw *llmservice.TextResult) string {
 	if raw == nil {
 		return ""
 	}
diff --git a/backend/cmd/start.go b/backend/cmd/start.go
index 5fca1ab..b8399fb 100644
--- a/backend/cmd/start.go
+++ b/backend/cmd/start.go
@@ -23,10 +23,7 @@ import (
 	"github.com/LoveLosita/smartflow/backend/api"
 	"github.com/LoveLosita/smartflow/backend/dao"
 	kafkabus "github.com/LoveLosita/smartflow/backend/infra/kafka"
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
 	outboxinfra "github.com/LoveLosita/smartflow/backend/infra/outbox"
-	infrarag "github.com/LoveLosita/smartflow/backend/infra/rag"
-	ragconfig "github.com/LoveLosita/smartflow/backend/infra/rag/config"
 	"github.com/LoveLosita/smartflow/backend/inits"
 	"github.com/LoveLosita/smartflow/backend/memory"
 	memorymodel "github.com/LoveLosita/smartflow/backend/memory/model"
@@ -44,6 +41,9 @@ import (
 	"github.com/LoveLosita/smartflow/backend/service"
 	agentsvcsvc "github.com/LoveLosita/smartflow/backend/service/agentsvc"
 	eventsvc "github.com/LoveLosita/smartflow/backend/service/events"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
+	ragservice "github.com/LoveLosita/smartflow/backend/services/rag"
+	ragconfig "github.com/LoveLosita/smartflow/backend/services/rag/config"
 	"github.com/go-redis/redis/v8"
 	"github.com/spf13/viper"
 	"gorm.io/gorm"
@@ -167,17 +167,25 @@ func buildRuntime(ctx context.Context) (*appRuntime, error) {
 		return nil, fmt.Errorf("failed to initialize Eino: %w", err)
 	}
 
-	ragRuntime, err := buildRAGRuntime(ctx)
+	llmService := llmservice.New(llmservice.Options{
+		AIHub:             aiHub,
+		APIKey:            os.Getenv("ARK_API_KEY"),
+		BaseURL:           viper.GetString("agent.baseURL"),
+		CourseVisionModel: viper.GetString("courseImport.visionModel"),
+	})
+
+	ragService, err := buildRAGService(ctx)
 	if err != nil {
 		return nil, err
 	}
+	ragRuntime := ragService.Runtime()
 
 	memoryCfg := memory.LoadConfigFromViper()
 	memoryObserver := memoryobserve.NewLoggerObserver(log.Default())
 	memoryMetrics := memoryobserve.NewMetricsRegistry()
 	memoryModule := memory.NewModuleWithObserve(
 		db,
-		infrallm.WrapArkClient(aiHub.Pro),
+		llmService.ProClient(),
 		ragRuntime,
 		memoryCfg,
 		memory.ObserveDeps{
@@ -208,11 +216,11 @@ func buildRuntime(ctx context.Context) (*appRuntime, error) {
 	userService := service.NewUserService(userRepo, cacheRepo)
 	taskSv := service.NewTaskService(taskRepo, cacheRepo, eventBus)
 	taskSv.SetActiveScheduleDAO(manager.ActiveSchedule)
-	courseService := buildCourseService(courseRepo, scheduleRepo)
+	courseService := buildCourseService(llmService, courseRepo, scheduleRepo)
 	taskClassService := service.NewTaskClassService(taskClassRepo, cacheRepo, scheduleRepo, manager)
 	scheduleService := service.NewScheduleService(scheduleRepo, userRepo, taskClassRepo, manager, cacheRepo)
 	agentService := service.NewAgentServiceWithSchedule(
-		aiHub,
+		llmService,
 		agentRepo,
 		taskRepo,
 		cacheRepo,
@@ -251,7 +259,7 @@ func buildRuntime(ctx context.Context) (*appRuntime, error) {
 	}
 	// 1. 主动调度选择器单独复用 Pro 模型，LLM 失败时由 selection 层显式回退到确定性候选；
 	// 2. dry-run 与 selection 通过 graph runner 串起来，避免 trigger_pipeline 再拼第二套候选逻辑。
-	activeScheduleLLMClient := infrallm.WrapArkClient(aiHub.Pro)
+	activeScheduleLLMClient := llmService.ProClient()
 	activeScheduleSelector := activesel.NewService(activeScheduleLLMClient)
 	activeScheduleFeedbackLocator := activefeedbacklocate.NewService(activeReaders, activeScheduleLLMClient)
 	activeScheduleGraphRunner, err := activegraph.NewRunner(activeScheduleDryRun.AsGraphDryRunFunc(), activeScheduleSelector)
@@ -323,26 +331,26 @@ func buildRuntime(ctx context.Context) (*appRuntime, error) {
 	return runtime, nil
 }
 
-func buildRAGRuntime(ctx context.Context) (infrarag.Runtime, error) {
+func buildRAGService(ctx context.Context) (*ragservice.Service, error) {
 	ragCfg := ragconfig.LoadFromViper()
 	if !ragCfg.Enabled {
-		log.Println("RAG runtime is disabled")
-		return nil, nil
+		log.Println("RAG service is disabled")
+		return ragservice.New(ragservice.Options{}), nil
 	}
 
 	// 1. 当前项目尚未完成全局观测平台建设，这里先注入一层轻量 Observer；
 	// 2. RAG 内部只依赖 Observer 接口，后续若全项目统一日志/指标系统，只需替换这里；
 	// 3. 这样可以避免 RAG 单独自建一套割裂的日志基础设施。
 	ragLogger := log.Default()
-	ragRuntime, err := infrarag.NewRuntimeFromConfig(ctx, ragCfg, infrarag.FactoryDeps{
+	ragService, err := ragservice.NewFromConfig(ctx, ragCfg, ragservice.FactoryDeps{
 		Logger:   ragLogger,
-		Observer: infrarag.NewLoggerObserver(ragLogger),
+		Observer: ragservice.NewLoggerObserver(ragLogger),
 	})
 	if err != nil {
-		return nil, fmt.Errorf("failed to initialize RAG runtime: %w", err)
+		return nil, fmt.Errorf("failed to initialize RAG service: %w", err)
 	}
-	log.Printf("RAG runtime initialized: store=%s embed=%s reranker=%s", ragCfg.Store, ragCfg.EmbedProvider, ragCfg.RerankerProvider)
-	return ragRuntime, nil
+	log.Printf("RAG service initialized: store=%s embed=%s reranker=%s", ragCfg.Store, ragCfg.EmbedProvider, ragCfg.RerankerProvider)
+	return ragService, nil
 }
 
 func buildEventBus(outboxRepo *outboxinfra.Repository) (eventsvc.OutboxBus, error) {
@@ -369,12 +377,8 @@ func buildEventBus(outboxRepo *outboxinfra.Repository) (eventsvc.OutboxBus, erro
 	return eventBus, nil
 }
 
-func buildCourseService(courseRepo *dao.CourseDAO, scheduleRepo *dao.ScheduleDAO) *service.CourseService {
-	courseImageResponsesClient := infrallm.NewArkResponsesClient(
-		os.Getenv("ARK_API_KEY"),
-		viper.GetString("agent.baseURL"),
-		viper.GetString("courseImport.visionModel"),
-	)
+func buildCourseService(llmService *llmservice.Service, courseRepo *dao.CourseDAO, scheduleRepo *dao.ScheduleDAO) *service.CourseService {
+	courseImageResponsesClient := llmService.CourseImageResponsesClient()
 	return service.NewCourseService(
 		courseRepo,
 		scheduleRepo,
@@ -650,7 +654,7 @@ func containsString(values []string, target string) bool {
 
 func configureAgentService(
 	agentService *service.AgentService,
-	ragRuntime infrarag.Runtime,
+	ragRuntime ragservice.Runtime,
 	agentRepo *dao.AgentDAO,
 	cacheRepo *dao.CacheDAO,
 	taskRepo *dao.TaskDAO,
diff --git a/backend/memory/module.go b/backend/memory/module.go
index 4e50414..a1deba0 100644
--- a/backend/memory/module.go
+++ b/backend/memory/module.go
@@ -5,8 +5,6 @@ import (
 	"errors"
 	"log"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
-	infrarag "github.com/LoveLosita/smartflow/backend/infra/rag"
 	memorycleanup "github.com/LoveLosita/smartflow/backend/memory/cleanup"
 	memorymodel "github.com/LoveLosita/smartflow/backend/memory/model"
 	memoryobserve "github.com/LoveLosita/smartflow/backend/memory/observe"
@@ -16,6 +14,8 @@ import (
 	memoryvectorsync "github.com/LoveLosita/smartflow/backend/memory/vectorsync"
 	memoryworker "github.com/LoveLosita/smartflow/backend/memory/worker"
 	"github.com/LoveLosita/smartflow/backend/model"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
+	ragservice "github.com/LoveLosita/smartflow/backend/services/rag"
 	"gorm.io/gorm"
 )
 
@@ -28,8 +28,8 @@ import (
 type Module struct {
 	db         *gorm.DB
 	cfg        memorymodel.Config
-	llmClient  *infrallm.Client
-	ragRuntime infrarag.Runtime
+	llmClient  *llmservice.Client
+	ragRuntime ragservice.Runtime
 	observer   memoryobserve.Observer
 	metrics    memoryobserve.MetricsRecorder
 
@@ -64,15 +64,15 @@ func LoadConfigFromViper() memorymodel.Config {
 // 2. llmClient 允许为 nil，此时写入链路会自动回退到本地 fallback 抽取；
 // 3. ragRuntime 允许为 nil，此时读取/向量同步自动回退旧逻辑；
 // 4. 若后续接入统一 DI 容器，也应优先注册这个 Module，而不是把内部 repo/service 继续向外泄漏。
-func NewModule(db *gorm.DB, llmClient *infrallm.Client, ragRuntime infrarag.Runtime, cfg memorymodel.Config) *Module {
+func NewModule(db *gorm.DB, llmClient *llmservice.Client, ragRuntime ragservice.Runtime, cfg memorymodel.Config) *Module {
 	return NewModuleWithObserve(db, llmClient, ragRuntime, cfg, ObserveDeps{})
 }
 
 // NewModuleWithObserve 创建带观测依赖的 memory 模块门面。
 func NewModuleWithObserve(
 	db *gorm.DB,
-	llmClient *infrallm.Client,
-	ragRuntime infrarag.Runtime,
+	llmClient *llmservice.Client,
+	ragRuntime ragservice.Runtime,
 	cfg memorymodel.Config,
 	deps ObserveDeps,
 ) *Module {
@@ -228,8 +228,8 @@ func (m *Module) StartWorker(ctx context.Context) {
 
 func wireModule(
 	db *gorm.DB,
-	llmClient *infrallm.Client,
-	ragRuntime infrarag.Runtime,
+	llmClient *llmservice.Client,
+	ragRuntime ragservice.Runtime,
 	cfg memorymodel.Config,
 	deps ObserveDeps,
 ) *Module {
diff --git a/backend/memory/orchestrator/llm_decision_orchestrator.go b/backend/memory/orchestrator/llm_decision_orchestrator.go
index 7f45da7..6769d5e 100644
--- a/backend/memory/orchestrator/llm_decision_orchestrator.go
+++ b/backend/memory/orchestrator/llm_decision_orchestrator.go
@@ -6,8 +6,8 @@ import (
 	"log"
 	"strings"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
 	memorymodel "github.com/LoveLosita/smartflow/backend/memory/model"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 )
 
 const defaultDecisionCompareMaxTokens = 600
@@ -19,13 +19,13 @@ const defaultDecisionCompareMaxTokens = 600
 // 2. LLM 只输出 relation（关系类型），不输出 action，不输出 target ID；
 // 3. LLM 调用失败时返回 error，由上层决定是否视为 unrelated。
 type LLMDecisionOrchestrator struct {
-	client *infrallm.Client
+	client *llmservice.Client
 	cfg    memorymodel.Config
 	logger *log.Logger
 }
 
 // NewLLMDecisionOrchestrator 构造决策比对编排器。
-func NewLLMDecisionOrchestrator(client *infrallm.Client, cfg memorymodel.Config) *LLMDecisionOrchestrator {
+func NewLLMDecisionOrchestrator(client *llmservice.Client, cfg memorymodel.Config) *LLMDecisionOrchestrator {
 	return &LLMDecisionOrchestrator{
 		client: client,
 		cfg:    cfg,
@@ -52,14 +52,14 @@ func (o *LLMDecisionOrchestrator) Compare(
 	systemPrompt := buildDecisionCompareSystemPrompt()
 	userPrompt := buildDecisionCompareUserPrompt(fact, candidate)
 
-	messages := infrallm.BuildSystemUserMessages(systemPrompt, nil, userPrompt)
+	messages := llmservice.BuildSystemUserMessages(systemPrompt, nil, userPrompt)
 
 	// 2. 调用 LLM 做结构化输出，温度用低值保证判断稳定。
-	resp, _, err := infrallm.GenerateJSON[decisionCompareResponse](
+	resp, _, err := llmservice.GenerateJSON[decisionCompareResponse](
 		ctx,
 		o.client,
 		messages,
-		infrallm.GenerateOptions{
+		llmservice.GenerateOptions{
 			Temperature: 0.1,
 			MaxTokens:   defaultDecisionCompareMaxTokens,
 			Thinking:    resolveMemoryThinkingMode(o.cfg.LLMThinking),
@@ -127,9 +127,9 @@ func buildDecisionCompareUserPrompt(fact memorymodel.NormalizedFact, candidate m
 }
 
 // resolveMemoryThinkingMode 根据配置布尔值返回对应的 ThinkingMode。
-func resolveMemoryThinkingMode(enabled bool) infrallm.ThinkingMode {
+func resolveMemoryThinkingMode(enabled bool) llmservice.ThinkingMode {
 	if enabled {
-		return infrallm.ThinkingModeEnabled
+		return llmservice.ThinkingModeEnabled
 	}
-	return infrallm.ThinkingModeDisabled
+	return llmservice.ThinkingModeDisabled
 }
diff --git a/backend/memory/orchestrator/llm_write_orchestrator.go b/backend/memory/orchestrator/llm_write_orchestrator.go
index ca15fa1..c5121a6 100644
--- a/backend/memory/orchestrator/llm_write_orchestrator.go
+++ b/backend/memory/orchestrator/llm_write_orchestrator.go
@@ -7,9 +7,9 @@ import (
 	"log"
 	"strings"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
 	memorymodel "github.com/LoveLosita/smartflow/backend/memory/model"
 	memoryutils "github.com/LoveLosita/smartflow/backend/memory/utils"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 )
 
 const (
@@ -24,13 +24,13 @@ const (
 // 2. 不负责落库，不负责任务状态机推进；
 // 3. 当 LLM 不可用或输出异常时，回退到保守的本地抽取，保证链路不完全断。
 type LLMWriteOrchestrator struct {
-	client *infrallm.Client
+	client *llmservice.Client
 	cfg    memorymodel.Config
 	logger *log.Logger
 }
 
 // NewLLMWriteOrchestrator 构造 LLM 版记忆写入编排器。
-func NewLLMWriteOrchestrator(client *infrallm.Client, cfg memorymodel.Config) *LLMWriteOrchestrator {
+func NewLLMWriteOrchestrator(client *llmservice.Client, cfg memorymodel.Config) *LLMWriteOrchestrator {
 	return &LLMWriteOrchestrator{
 		client: client,
 		cfg:    cfg,
@@ -54,17 +54,17 @@ func (o *LLMWriteOrchestrator) ExtractFacts(ctx context.Context, payload memorym
 		return fallbackNormalizedFacts(payload), nil
 	}
 
-	messages := infrallm.BuildSystemUserMessages(
+	messages := llmservice.BuildSystemUserMessages(
 		buildMemoryExtractSystemPrompt(o.cfg.ExtractPrompt),
 		nil,
 		buildMemoryExtractUserPrompt(payload),
 	)
 
-	resp, rawResult, err := infrallm.GenerateJSON[memoryExtractResponse](
+	resp, rawResult, err := llmservice.GenerateJSON[memoryExtractResponse](
 		ctx,
 		o.client,
 		messages,
-		infrallm.GenerateOptions{
+		llmservice.GenerateOptions{
 			Temperature: clampTemperature(o.cfg.LLMTemperature),
 			MaxTokens:   defaultMemoryExtractMaxTokens,
 			Thinking:    resolveMemoryThinkingMode(o.cfg.LLMThinking),
@@ -319,7 +319,7 @@ func isSkipIntent(intent string) bool {
 	}
 }
 
-func truncateForLog(raw *infrallm.TextResult) string {
+func truncateForLog(raw *llmservice.TextResult) string {
 	if raw == nil {
 		return ""
 	}
diff --git a/backend/memory/service/read_scope.go b/backend/memory/service/read_scope.go
index ed8c376..a4c03a1 100644
--- a/backend/memory/service/read_scope.go
+++ b/backend/memory/service/read_scope.go
@@ -3,8 +3,8 @@ package service
 import (
 	"time"
 
-	infrarag "github.com/LoveLosita/smartflow/backend/infra/rag"
 	memorymodel "github.com/LoveLosita/smartflow/backend/memory/model"
+	ragservice "github.com/LoveLosita/smartflow/backend/services/rag"
 )
 
 // buildReadScopedItemQuery 构造读侧统一使用的 MySQL 查询条件。
@@ -53,8 +53,8 @@ func buildReadScopedRAGRequest(
 	req memorymodel.RetrieveRequest,
 	topK int,
 	threshold float64,
-) infrarag.MemoryRetrieveRequest {
-	return infrarag.MemoryRetrieveRequest{
+) ragservice.MemoryRetrieveRequest {
+	return ragservice.MemoryRetrieveRequest{
 		Query:          req.Query,
 		TopK:           topK,
 		Threshold:      threshold,
diff --git a/backend/memory/service/read_service.go b/backend/memory/service/read_service.go
index 4302057..158c84e 100644
--- a/backend/memory/service/read_service.go
+++ b/backend/memory/service/read_service.go
@@ -8,12 +8,12 @@ import (
 	"strings"
 	"time"
 
-	infrarag "github.com/LoveLosita/smartflow/backend/infra/rag"
 	memorymodel "github.com/LoveLosita/smartflow/backend/memory/model"
 	memoryobserve "github.com/LoveLosita/smartflow/backend/memory/observe"
 	memoryrepo "github.com/LoveLosita/smartflow/backend/memory/repo"
 	memoryutils "github.com/LoveLosita/smartflow/backend/memory/utils"
 	"github.com/LoveLosita/smartflow/backend/model"
+	ragservice "github.com/LoveLosita/smartflow/backend/services/rag"
 )
 
 const (
@@ -30,7 +30,7 @@ const (
 type ReadService struct {
 	itemRepo     *memoryrepo.ItemRepo
 	settingsRepo *memoryrepo.SettingsRepo
-	ragRuntime   infrarag.Runtime
+	ragRuntime   ragservice.Runtime
 	cfg          memorymodel.Config
 	observer     memoryobserve.Observer
 	metrics      memoryobserve.MetricsRecorder
@@ -57,7 +57,7 @@ type semanticRetrieveTelemetry struct {
 func NewReadService(
 	itemRepo *memoryrepo.ItemRepo,
 	settingsRepo *memoryrepo.SettingsRepo,
-	ragRuntime infrarag.Runtime,
+	ragRuntime ragservice.Runtime,
 	cfg memorymodel.Config,
 	observer memoryobserve.Observer,
 	metrics memoryobserve.MetricsRecorder,
@@ -347,7 +347,7 @@ func collectMemoryIDs(items []model.MemoryItem) []int64 {
 	return ids
 }
 
-func buildMemoryDTOFromRetrieveHit(hit infrarag.RetrieveHit) (memorymodel.ItemDTO, int64) {
+func buildMemoryDTOFromRetrieveHit(hit ragservice.RetrieveHit) (memorymodel.ItemDTO, int64) {
 	memoryID := parseMemoryIDFromDocumentID(hit.DocumentID)
 	metadata := hit.Metadata
 	content := strings.TrimSpace(hit.Text)
diff --git a/backend/memory/vectorsync/syncer.go b/backend/memory/vectorsync/syncer.go
index 29bce7d..91b56ed 100644
--- a/backend/memory/vectorsync/syncer.go
+++ b/backend/memory/vectorsync/syncer.go
@@ -6,10 +6,10 @@ import (
 	"log"
 	"strings"
 
-	infrarag "github.com/LoveLosita/smartflow/backend/infra/rag"
 	memoryobserve "github.com/LoveLosita/smartflow/backend/memory/observe"
 	memoryrepo "github.com/LoveLosita/smartflow/backend/memory/repo"
 	"github.com/LoveLosita/smartflow/backend/model"
+	ragservice "github.com/LoveLosita/smartflow/backend/services/rag"
 )
 
 // Syncer 负责 memory_items 与向量库之间的最小桥接。
@@ -19,7 +19,7 @@ import (
 // 2. 不负责决定哪些记忆该写、该删、该恢复，这些决策仍由上游 service/worker/cleanup 控制；
 // 3. 同步失败时只回写 vector_status 并打观测，不反向回滚业务事务，避免把在线链路拖成强依赖。
 type Syncer struct {
-	ragRuntime infrarag.Runtime
+	ragRuntime ragservice.Runtime
 	itemRepo   *memoryrepo.ItemRepo
 	observer   memoryobserve.Observer
 	metrics    memoryobserve.MetricsRecorder
@@ -27,7 +27,7 @@ type Syncer struct {
 }
 
 func NewSyncer(
-	ragRuntime infrarag.Runtime,
+	ragRuntime ragservice.Runtime,
 	itemRepo *memoryrepo.ItemRepo,
 	observer memoryobserve.Observer,
 	metrics memoryobserve.MetricsRecorder,
@@ -53,9 +53,9 @@ func (s *Syncer) Upsert(ctx context.Context, traceID string, items []model.Memor
 		return
 	}
 
-	requestItems := make([]infrarag.MemoryIngestItem, 0, len(items))
+	requestItems := make([]ragservice.MemoryIngestItem, 0, len(items))
 	for _, item := range items {
-		requestItems = append(requestItems, infrarag.MemoryIngestItem{
+		requestItems = append(requestItems, ragservice.MemoryIngestItem{
 			MemoryID:         item.ID,
 			UserID:           item.UserID,
 			ConversationID:   strValue(item.ConversationID),
@@ -76,7 +76,7 @@ func (s *Syncer) Upsert(ctx context.Context, traceID string, items []model.Memor
 
 	result, err := s.ragRuntime.IngestMemory(memoryobserve.WithFields(ctx, map[string]any{
 		"trace_id": traceID,
-	}), infrarag.MemoryIngestRequest{
+	}), ragservice.MemoryIngestRequest{
 		TraceID: traceID,
 		Action:  "add",
 		Items:   requestItems,
diff --git a/backend/memory/worker/decision_flow.go b/backend/memory/worker/decision_flow.go
index 9be964a..e2d080d 100644
--- a/backend/memory/worker/decision_flow.go
+++ b/backend/memory/worker/decision_flow.go
@@ -4,11 +4,11 @@ import (
 	"context"
 	"fmt"
 
-	infrarag "github.com/LoveLosita/smartflow/backend/infra/rag"
 	memorymodel "github.com/LoveLosita/smartflow/backend/memory/model"
 	memoryrepo "github.com/LoveLosita/smartflow/backend/memory/repo"
 	memoryutils "github.com/LoveLosita/smartflow/backend/memory/utils"
 	"github.com/LoveLosita/smartflow/backend/model"
+	ragservice "github.com/LoveLosita/smartflow/backend/services/rag"
 	"gorm.io/gorm"
 )
 
@@ -192,7 +192,7 @@ func (r *Runner) recallCandidates(
 ) candidateRecallResult {
 	// 1. 优先使用 Milvus 向量语义召回。
 	if r.ragRuntime != nil {
-		retrieveResult, err := r.ragRuntime.RetrieveMemory(ctx, infrarag.MemoryRetrieveRequest{
+		retrieveResult, err := r.ragRuntime.RetrieveMemory(ctx, ragservice.MemoryRetrieveRequest{
 			Query:       fact.Content,
 			TopK:        r.cfg.DecisionCandidateTopK,
 			Threshold:   r.cfg.DecisionCandidateMinScore,
@@ -235,7 +235,7 @@ func (r *Runner) recallCandidates(
 // 1. 从 DocumentID（格式 memory:{id}）解析出 mysql_id；
 // 2. 从 metadata 提取 title 和 memory_type；
 // 3. 跳过无法解析 DocumentID 的结果。
-func (r *Runner) buildCandidatesFromRAG(hits []infrarag.RetrieveHit) []memorymodel.CandidateSnapshot {
+func (r *Runner) buildCandidatesFromRAG(hits []ragservice.RetrieveHit) []memorymodel.CandidateSnapshot {
 	candidates := make([]memorymodel.CandidateSnapshot, 0, len(hits))
 	for _, hit := range hits {
 		memoryID := parseMemoryID(hit.DocumentID)
diff --git a/backend/memory/worker/runner.go b/backend/memory/worker/runner.go
index 650ea0a..bd49bb2 100644
--- a/backend/memory/worker/runner.go
+++ b/backend/memory/worker/runner.go
@@ -9,7 +9,6 @@ import (
 	"strings"
 	"time"
 
-	infrarag "github.com/LoveLosita/smartflow/backend/infra/rag"
 	memorymodel "github.com/LoveLosita/smartflow/backend/memory/model"
 	memoryobserve "github.com/LoveLosita/smartflow/backend/memory/observe"
 	memoryorchestrator "github.com/LoveLosita/smartflow/backend/memory/orchestrator"
@@ -17,6 +16,7 @@ import (
 	memoryutils "github.com/LoveLosita/smartflow/backend/memory/utils"
 	memoryvectorsync "github.com/LoveLosita/smartflow/backend/memory/vectorsync"
 	"github.com/LoveLosita/smartflow/backend/model"
+	ragservice "github.com/LoveLosita/smartflow/backend/services/rag"
 	"gorm.io/gorm"
 )
 
@@ -41,7 +41,7 @@ type Runner struct {
 	auditRepo    *memoryrepo.AuditRepo
 	settingsRepo *memoryrepo.SettingsRepo
 	extractor    Extractor
-	ragRuntime   infrarag.Runtime
+	ragRuntime   ragservice.Runtime
 	logger       *log.Logger
 	vectorSyncer *memoryvectorsync.Syncer
 	observer     memoryobserve.Observer
@@ -63,7 +63,7 @@ func NewRunner(
 	auditRepo *memoryrepo.AuditRepo,
 	settingsRepo *memoryrepo.SettingsRepo,
 	extractor Extractor,
-	ragRuntime infrarag.Runtime,
+	ragRuntime ragservice.Runtime,
 	cfg memorymodel.Config,
 	decisionOrchestrator *memoryorchestrator.LLMDecisionOrchestrator,
 	vectorSyncer *memoryvectorsync.Syncer,
diff --git a/backend/newAgent/model/graph_run_state.go b/backend/newAgent/model/graph_run_state.go
index 7193fae..52d0056 100644
--- a/backend/newAgent/model/graph_run_state.go
+++ b/backend/newAgent/model/graph_run_state.go
@@ -5,10 +5,10 @@ import (
 	"strings"
 	"time"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
 	newagentstream "github.com/LoveLosita/smartflow/backend/newAgent/stream"
 	newagenttools "github.com/LoveLosita/smartflow/backend/newAgent/tools"
 	schedule "github.com/LoveLosita/smartflow/backend/newAgent/tools/schedule"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 	"github.com/cloudwego/eino/schema"
 )
 
@@ -71,10 +71,10 @@ type PersistVisibleMessageFunc func(ctx context.Context, state *CommonState, msg
 // 2. Chat/Plan/Execute/Deliver 允许分别挂不同 client，但也允许先复用同一个 client；
 // 3. ChunkEmitter 统一承接阶段提示、正文、工具事件、确认请求等 SSE 输出。
 type AgentGraphDeps struct {
-	ChatClient           *infrallm.Client
-	PlanClient           *infrallm.Client
-	ExecuteClient        *infrallm.Client
-	DeliverClient        *infrallm.Client
+	ChatClient           *llmservice.Client
+	PlanClient           *llmservice.Client
+	ExecuteClient        *llmservice.Client
+	DeliverClient        *llmservice.Client
 	ChunkEmitter         *newagentstream.ChunkEmitter
 	StateStore           AgentStateStore
 	ToolRegistry         *newagenttools.ToolRegistry
@@ -141,7 +141,7 @@ func (d *AgentGraphDeps) EnsureChunkEmitter() *newagentstream.ChunkEmitter {
 }
 
 // ResolveChatClient 返回 chat 阶段可用的模型客户端。
-func (d *AgentGraphDeps) ResolveChatClient() *infrallm.Client {
+func (d *AgentGraphDeps) ResolveChatClient() *llmservice.Client {
 	if d == nil {
 		return nil
 	}
@@ -154,7 +154,7 @@ func (d *AgentGraphDeps) ResolveChatClient() *infrallm.Client {
 // 1. 优先使用显式注入的 PlanClient；
 // 2. 若未单独注入，则回退到 ChatClient；
 // 3. 这样在骨架期可先用一套 client 跑通，再按需拆分 strategist / worker。
-func (d *AgentGraphDeps) ResolvePlanClient() *infrallm.Client {
+func (d *AgentGraphDeps) ResolvePlanClient() *llmservice.Client {
 	if d == nil {
 		return nil
 	}
@@ -165,7 +165,7 @@ func (d *AgentGraphDeps) ResolvePlanClient() *infrallm.Client {
 }
 
 // ResolveExecuteClient 返回 execute 阶段可用的模型客户端。
-func (d *AgentGraphDeps) ResolveExecuteClient() *infrallm.Client {
+func (d *AgentGraphDeps) ResolveExecuteClient() *llmservice.Client {
 	if d == nil {
 		return nil
 	}
@@ -179,7 +179,7 @@ func (d *AgentGraphDeps) ResolveExecuteClient() *infrallm.Client {
 }
 
 // ResolveDeliverClient 返回 deliver 阶段可用的模型客户端。
-func (d *AgentGraphDeps) ResolveDeliverClient() *infrallm.Client {
+func (d *AgentGraphDeps) ResolveDeliverClient() *llmservice.Client {
 	if d == nil {
 		return nil
 	}
diff --git a/backend/newAgent/node/chat.go b/backend/newAgent/node/chat.go
index 24876bc..aa3d63b 100644
--- a/backend/newAgent/node/chat.go
+++ b/backend/newAgent/node/chat.go
@@ -11,11 +11,11 @@ import (
 	"github.com/cloudwego/eino/schema"
 	"github.com/google/uuid"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
 	newagentmodel "github.com/LoveLosita/smartflow/backend/newAgent/model"
 	newagentprompt "github.com/LoveLosita/smartflow/backend/newAgent/prompt"
 	newagentrouter "github.com/LoveLosita/smartflow/backend/newAgent/router"
 	newagentstream "github.com/LoveLosita/smartflow/backend/newAgent/stream"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 )
 
 const (
@@ -50,7 +50,7 @@ type ChatNodeInput struct {
 	UserInput             string
 	ConfirmAction         string
 	ResumeInteractionID   string
-	Client                *infrallm.Client
+	Client                *llmservice.Client
 	ChunkEmitter          *newagentstream.ChunkEmitter
 	CompactionStore       newagentmodel.CompactionStore // 上下文压缩持久化
 	PersistVisibleMessage newagentmodel.PersistVisibleMessageFunc
@@ -107,9 +107,9 @@ func RunChatNode(ctx context.Context, input ChatNodeInput) error {
 	})
 	logNodeLLMContext(chatStageName, "routing", flowState, messages)
 
-	reader, err := input.Client.Stream(ctx, messages, infrallm.GenerateOptions{
+	reader, err := input.Client.Stream(ctx, messages, llmservice.GenerateOptions{
 		Temperature: 0.7,
-		Thinking:    infrallm.ThinkingModeDisabled,
+		Thinking:    llmservice.ThinkingModeDisabled,
 		Metadata: map[string]any{
 			"stage": chatStageName,
 			"phase": "routing",
@@ -172,7 +172,7 @@ func isExecuteLoopClosedMarker(msg *schema.Message) bool {
 // 3. 控制码解析超时或流异常结束 → fallback 到 plan。
 func streamAndDispatch(
 	ctx context.Context,
-	reader infrallm.StreamReader,
+	reader llmservice.StreamReader,
 	parser *newagentrouter.StreamRouteParser,
 	input ChatNodeInput,
 	emitter *newagentstream.ChunkEmitter,
@@ -292,7 +292,7 @@ func resolveEffectiveThinking(mode string, route newagentmodel.ChatRoute, decisi
 // 2. thinking=true：关闭路由流，发起第二次 thinking 流式调用。
 func handleDirectReplyStream(
 	ctx context.Context,
-	reader infrallm.StreamReader,
+	reader llmservice.StreamReader,
 	input ChatNodeInput,
 	emitter *newagentstream.ChunkEmitter,
 	conversationContext *newagentmodel.ConversationContext,
@@ -309,7 +309,7 @@ func handleDirectReplyStream(
 // handleThinkingReplyStream 处理需要思考的回复：关闭路由流 → 第二次 thinking 流式调用。
 func handleThinkingReplyStream(
 	ctx context.Context,
-	reader infrallm.StreamReader,
+	reader llmservice.StreamReader,
 	input ChatNodeInput,
 	emitter *newagentstream.ChunkEmitter,
 	conversationContext *newagentmodel.ConversationContext,
@@ -327,10 +327,10 @@ func handleThinkingReplyStream(
 		StatusBlockID:   chatStatusBlockID,
 	})
 	logNodeLLMContext(chatStageName, "direct_reply_thinking", flowState, deepMessages)
-	deepReader, err := input.Client.Stream(ctx, deepMessages, infrallm.GenerateOptions{
+	deepReader, err := input.Client.Stream(ctx, deepMessages, llmservice.GenerateOptions{
 		Temperature: 0.5,
 		MaxTokens:   2000,
-		Thinking:    infrallm.ThinkingModeEnabled,
+		Thinking:    llmservice.ThinkingModeEnabled,
 		Metadata: map[string]any{
 			"stage": chatStageName,
 			"phase": "direct_reply_thinking",
@@ -363,7 +363,7 @@ func handleThinkingReplyStream(
 // handleDirectReplyContinueStream 处理无思考的闲聊：同一流续传。
 func handleDirectReplyContinueStream(
 	ctx context.Context,
-	reader infrallm.StreamReader,
+	reader llmservice.StreamReader,
 	input ChatNodeInput,
 	emitter *newagentstream.ChunkEmitter,
 	conversationContext *newagentmodel.ConversationContext,
@@ -419,7 +419,7 @@ func handleDirectReplyContinueStream(
 // 2. 推送轻量状态通知；
 // 3. 设置流程状态，进入 Execute 或 RoughBuild。
 func handleRouteExecuteStream(
-	reader infrallm.StreamReader,
+	reader llmservice.StreamReader,
 	emitter *newagentstream.ChunkEmitter,
 	flowState *newagentmodel.CommonState,
 	decision *newagentmodel.ChatRoutingDecision,
@@ -674,7 +674,7 @@ func isExplicitNoRefineAfterRoughBuildRequest(userInput string) bool {
 // 4. 完整回复写入 history。
 func handleDeepAnswerStream(
 	ctx context.Context,
-	reader infrallm.StreamReader,
+	reader llmservice.StreamReader,
 	input ChatNodeInput,
 	emitter *newagentstream.ChunkEmitter,
 	conversationContext *newagentmodel.ConversationContext,
@@ -685,9 +685,9 @@ func handleDeepAnswerStream(
 	_ = reader.Close()
 
 	// 2. 第二次流式调用。
-	thinkingOpt := infrallm.ThinkingModeDisabled
+	thinkingOpt := llmservice.ThinkingModeDisabled
 	if effectiveThinking {
-		thinkingOpt = infrallm.ThinkingModeEnabled
+		thinkingOpt = llmservice.ThinkingModeEnabled
 	}
 	deepMessages := newagentprompt.BuildDeepAnswerMessages(flowState, conversationContext, input.UserInput)
 	deepMessages = compactUnifiedMessagesIfNeeded(ctx, deepMessages, UnifiedCompactInput{
@@ -699,7 +699,7 @@ func handleDeepAnswerStream(
 		StatusBlockID:   chatStatusBlockID,
 	})
 	logNodeLLMContext(chatStageName, "deep_answer", flowState, deepMessages)
-	deepReader, err := input.Client.Stream(ctx, deepMessages, infrallm.GenerateOptions{
+	deepReader, err := input.Client.Stream(ctx, deepMessages, llmservice.GenerateOptions{
 		Temperature: 0.5,
 		MaxTokens:   2000,
 		Thinking:    thinkingOpt,
@@ -741,7 +741,7 @@ func handleDeepAnswerStream(
 
 // handleRoutePlanStream 处理规划路由：推送状态确认 → 设 PhasePlanning。
 func handleRoutePlanStream(
-	reader infrallm.StreamReader,
+	reader llmservice.StreamReader,
 	emitter *newagentstream.ChunkEmitter,
 	flowState *newagentmodel.CommonState,
 	effectiveThinking bool,
diff --git a/backend/newAgent/node/deliver.go b/backend/newAgent/node/deliver.go
index 6e0da3a..3082c1a 100644
--- a/backend/newAgent/node/deliver.go
+++ b/backend/newAgent/node/deliver.go
@@ -9,10 +9,10 @@ import (
 
 	"github.com/cloudwego/eino/schema"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
 	newagentmodel "github.com/LoveLosita/smartflow/backend/newAgent/model"
 	newagentprompt "github.com/LoveLosita/smartflow/backend/newAgent/prompt"
 	newagentstream "github.com/LoveLosita/smartflow/backend/newAgent/stream"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 )
 
 const (
@@ -31,7 +31,7 @@ const (
 type DeliverNodeInput struct {
 	RuntimeState          *newagentmodel.AgentRuntimeState
 	ConversationContext   *newagentmodel.ConversationContext
-	Client                *infrallm.Client
+	Client                *llmservice.Client
 	ChunkEmitter          *newagentstream.ChunkEmitter
 	ThinkingEnabled       bool                          // 是否开启 thinking，由 config.yaml 的 agent.thinking.deliver 注入
 	CompactionStore       newagentmodel.CompactionStore // 上下文压缩持久化
@@ -128,7 +128,7 @@ func RunDeliverNode(ctx context.Context, input DeliverNodeInput) error {
 //   - streamed：true 表示文本已通过 EmitStreamAssistantText 真流式推送到前端，调用方无需再伪流式。
 func generateDeliverSummary(
 	ctx context.Context,
-	client *infrallm.Client,
+	client *llmservice.Client,
 	flowState *newagentmodel.CommonState,
 	conversationContext *newagentmodel.ConversationContext,
 	thinkingEnabled bool,
@@ -162,7 +162,7 @@ func generateDeliverSummary(
 	reader, err := client.Stream(
 		ctx,
 		messages,
-		infrallm.GenerateOptions{
+		llmservice.GenerateOptions{
 			Temperature: 0.5,
 			MaxTokens:   800,
 			Thinking:    resolveThinkingMode(thinkingEnabled),
diff --git a/backend/newAgent/node/execute/action_router.go b/backend/newAgent/node/execute/action_router.go
index d628a58..399a358 100644
--- a/backend/newAgent/node/execute/action_router.go
+++ b/backend/newAgent/node/execute/action_router.go
@@ -8,11 +8,11 @@ import (
 	"log"
 	"strings"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
 	newagentmodel "github.com/LoveLosita/smartflow/backend/newAgent/model"
 	newagentrouter "github.com/LoveLosita/smartflow/backend/newAgent/router"
 	newagentstream "github.com/LoveLosita/smartflow/backend/newAgent/stream"
 	newagenttools "github.com/LoveLosita/smartflow/backend/newAgent/tools"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 	"github.com/cloudwego/eino/schema"
 	"github.com/google/uuid"
 )
@@ -38,7 +38,7 @@ func collectExecuteDecisionFromLLM(
 	reader, err := input.Client.Stream(
 		ctx,
 		messages,
-		infrallm.GenerateOptions{
+		llmservice.GenerateOptions{
 			Temperature: 1.0,
 			MaxTokens:   131072,
 			Thinking:    newagentshared.ResolveThinkingMode(input.ThinkingEnabled),
@@ -123,7 +123,7 @@ func collectExecuteDecisionFromLLM(
 			return nil, nil
 		}
 
-		decision, parseErr := infrallm.ParseJSONObject[newagentmodel.ExecuteDecision](result.DecisionJSON)
+		decision, parseErr := llmservice.ParseJSONObject[newagentmodel.ExecuteDecision](result.DecisionJSON)
 		if parseErr != nil {
 			log.Printf(
 				"[DEBUG] execute LLM JSON 解析失败 chat=%s round=%d json=%s raw=%s",
diff --git a/backend/newAgent/node/execute/run.go b/backend/newAgent/node/execute/run.go
index ca91961..0b74e2b 100644
--- a/backend/newAgent/node/execute/run.go
+++ b/backend/newAgent/node/execute/run.go
@@ -5,12 +5,12 @@ import (
 	"fmt"
 	newagentshared "github.com/LoveLosita/smartflow/backend/newAgent/shared"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
 	newagentmodel "github.com/LoveLosita/smartflow/backend/newAgent/model"
 	newagentprompt "github.com/LoveLosita/smartflow/backend/newAgent/prompt"
 	newagentstream "github.com/LoveLosita/smartflow/backend/newAgent/stream"
 	newagenttools "github.com/LoveLosita/smartflow/backend/newAgent/tools"
 	"github.com/LoveLosita/smartflow/backend/newAgent/tools/schedule"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 )
 
 const (
@@ -29,7 +29,7 @@ type ExecuteNodeInput struct {
 	RuntimeState          *newagentmodel.AgentRuntimeState
 	ConversationContext   *newagentmodel.ConversationContext
 	UserInput             string
-	Client                *infrallm.Client
+	Client                *llmservice.Client
 	ChunkEmitter          *newagentstream.ChunkEmitter
 	ResumeNode            string
 	ToolRegistry          *newagenttools.ToolRegistry
diff --git a/backend/newAgent/node/plan.go b/backend/newAgent/node/plan.go
index aaf8e8b..6d8b2dc 100644
--- a/backend/newAgent/node/plan.go
+++ b/backend/newAgent/node/plan.go
@@ -10,11 +10,11 @@ import (
 
 	"github.com/google/uuid"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
 	newagentmodel "github.com/LoveLosita/smartflow/backend/newAgent/model"
 	newagentprompt "github.com/LoveLosita/smartflow/backend/newAgent/prompt"
 	newagentrouter "github.com/LoveLosita/smartflow/backend/newAgent/router"
 	newagentstream "github.com/LoveLosita/smartflow/backend/newAgent/stream"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 	"github.com/cloudwego/eino/schema"
 )
 
@@ -34,7 +34,7 @@ type PlanNodeInput struct {
 	RuntimeState          *newagentmodel.AgentRuntimeState
 	ConversationContext   *newagentmodel.ConversationContext
 	UserInput             string
-	Client                *infrallm.Client
+	Client                *llmservice.Client
 	ChunkEmitter          *newagentstream.ChunkEmitter
 	ResumeNode            string
 	AlwaysExecute         bool                          // true 时计划生成后自动确认，不进入 confirm 节点
@@ -87,7 +87,7 @@ func RunPlanNode(ctx context.Context, input PlanNodeInput) error {
 	reader, err := input.Client.Stream(
 		ctx,
 		messages,
-		infrallm.GenerateOptions{
+		llmservice.GenerateOptions{
 			Temperature: 0.2,
 			// 显式设置上限，避免依赖框架默认值（默认 4096）导致长决策被截断。
 			// 注意：当前模型接口 max_tokens 上限为 131072，超过会 400。
@@ -149,7 +149,7 @@ func RunPlanNode(ctx context.Context, input PlanNodeInput) error {
 			return fmt.Errorf("规划解析失败，原始输出=%s", result.RawBuffer)
 		}
 
-		decision, parseErr := infrallm.ParseJSONObject[newagentmodel.PlanDecision](result.DecisionJSON)
+		decision, parseErr := llmservice.ParseJSONObject[newagentmodel.PlanDecision](result.DecisionJSON)
 		if parseErr != nil {
 			return fmt.Errorf("规划决策 JSON 解析失败: %w (raw=%s)", parseErr, result.RawBuffer)
 		}
@@ -390,9 +390,9 @@ func buildPinnedPlanText(steps []newagentmodel.PlanStep) string {
 
 // resolveThinkingMode 根据配置布尔值返回对应的 ThinkingMode。
 // 供 plan / execute / deliver 节点统一使用。
-func resolveThinkingMode(enabled bool) infrallm.ThinkingMode {
+func resolveThinkingMode(enabled bool) llmservice.ThinkingMode {
 	if enabled {
-		return infrallm.ThinkingModeEnabled
+		return llmservice.ThinkingModeEnabled
 	}
-	return infrallm.ThinkingModeDisabled
+	return llmservice.ThinkingModeDisabled
 }
diff --git a/backend/newAgent/node/quick_task.go b/backend/newAgent/node/quick_task.go
index 82e0d38..5a4d405 100644
--- a/backend/newAgent/node/quick_task.go
+++ b/backend/newAgent/node/quick_task.go
@@ -8,13 +8,13 @@ import (
 	"strings"
 	"time"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
 	taskmodel "github.com/LoveLosita/smartflow/backend/model"
 	newagentmodel "github.com/LoveLosita/smartflow/backend/newAgent/model"
 	newagentprompt "github.com/LoveLosita/smartflow/backend/newAgent/prompt"
 	newagentrouter "github.com/LoveLosita/smartflow/backend/newAgent/router"
 	newagentshared "github.com/LoveLosita/smartflow/backend/newAgent/shared"
 	newagentstream "github.com/LoveLosita/smartflow/backend/newAgent/stream"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 	"github.com/cloudwego/eino/schema"
 )
 
@@ -30,7 +30,7 @@ type QuickTaskNodeInput struct {
 	RuntimeState          *newagentmodel.AgentRuntimeState
 	ConversationContext   *newagentmodel.ConversationContext
 	UserInput             string
-	Client                *infrallm.Client
+	Client                *llmservice.Client
 	ChunkEmitter          *newagentstream.ChunkEmitter
 	QuickTaskDeps         newagentmodel.QuickTaskDeps
 	PersistVisibleMessage newagentmodel.PersistVisibleMessageFunc
@@ -77,7 +77,7 @@ func RunQuickTaskNode(ctx context.Context, input QuickTaskNodeInput) error {
 	messages := newagentprompt.BuildQuickTaskMessagesSimple(input.UserInput)
 
 	// 2. 真流式调用 LLM。
-	reader, err := input.Client.Stream(ctx, messages, infrallm.GenerateOptions{
+	reader, err := input.Client.Stream(ctx, messages, llmservice.GenerateOptions{
 		Temperature: 0.3,
 		MaxTokens:   512,
 	})
@@ -130,7 +130,7 @@ func RunQuickTaskNode(ctx context.Context, input QuickTaskNodeInput) error {
 		// 解析 JSON。
 		log.Printf("[DEBUG] quick_task: LLM 原始决策 JSON chat=%s json=%s", flowState.ConversationID, result.DecisionJSON)
 		var parseErr error
-		decision, parseErr = infrallm.ParseJSONObject[quickTaskDecision](result.DecisionJSON)
+		decision, parseErr = llmservice.ParseJSONObject[quickTaskDecision](result.DecisionJSON)
 		if parseErr != nil {
 			log.Printf("[DEBUG] quick_task: JSON 解析失败 chat=%s json=%s", flowState.ConversationID, result.DecisionJSON)
 			if result.RawBuffer != "" {
diff --git a/backend/newAgent/node/unified_compact.go b/backend/newAgent/node/unified_compact.go
index bce1eb6..91cad2a 100644
--- a/backend/newAgent/node/unified_compact.go
+++ b/backend/newAgent/node/unified_compact.go
@@ -6,11 +6,11 @@ import (
 	"fmt"
 	"log"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
 	newagentmodel "github.com/LoveLosita/smartflow/backend/newAgent/model"
 	newagentprompt "github.com/LoveLosita/smartflow/backend/newAgent/prompt"
 	newagentstream "github.com/LoveLosita/smartflow/backend/newAgent/stream"
 	"github.com/LoveLosita/smartflow/backend/pkg"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 	"github.com/cloudwego/eino/schema"
 )
 
@@ -22,7 +22,7 @@ import (
 //  3. StageName 和 StatusBlockID 用于区分日志来源和 SSE 状态推送。
 type UnifiedCompactInput struct {
 	// Client 用于调用 LLM 压缩 msg1/msg2。
-	Client *infrallm.Client
+	Client *llmservice.Client
 	// CompactionStore 用于持久化压缩摘要和 token 统计，为 nil 时跳过持久化。
 	CompactionStore newagentmodel.CompactionStore
 	// FlowState 提供 userID / chatID / roundUsed 等定位信息。
diff --git a/backend/newAgent/prompt/compact_msg1.go b/backend/newAgent/prompt/compact_msg1.go
index 2fd2c61..6e8aed5 100644
--- a/backend/newAgent/prompt/compact_msg1.go
+++ b/backend/newAgent/prompt/compact_msg1.go
@@ -4,7 +4,7 @@ import (
 	"context"
 	"fmt"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 	"github.com/cloudwego/eino/schema"
 )
 
@@ -24,7 +24,7 @@ const compactMsg1SystemPrompt = `你是一个对话压缩助手。你的任务
 // existingSummary 不为空时表示已有旧摘要，需要合并压缩。
 func CompactMsg1(
 	ctx context.Context,
-	client *infrallm.Client,
+	client *llmservice.Client,
 	historyText string,
 	existingSummary string,
 ) (string, error) {
@@ -49,7 +49,7 @@ func CompactMsg1(
 		schema.UserMessage(userContent),
 	}
 
-	result, err := client.GenerateText(ctx, messages, infrallm.GenerateOptions{
+	result, err := client.GenerateText(ctx, messages, llmservice.GenerateOptions{
 		MaxTokens: 4000,
 	})
 	if err != nil {
diff --git a/backend/newAgent/prompt/compact_msg2.go b/backend/newAgent/prompt/compact_msg2.go
index b7ebb43..95605c5 100644
--- a/backend/newAgent/prompt/compact_msg2.go
+++ b/backend/newAgent/prompt/compact_msg2.go
@@ -4,7 +4,7 @@ import (
 	"context"
 	"fmt"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 	"github.com/cloudwego/eino/schema"
 )
 
@@ -23,7 +23,7 @@ const compactMsg2SystemPrompt = `你是一个执行记录压缩助手。你的
 // recentText 是保留的近期记录原文，不参与压缩。
 func CompactMsg2(
 	ctx context.Context,
-	client *infrallm.Client,
+	client *llmservice.Client,
 	earlyLoopText string,
 ) (string, error) {
 	userContent := fmt.Sprintf(`早期的 ReAct 执行记录：
@@ -36,7 +36,7 @@ func CompactMsg2(
 		schema.UserMessage(userContent),
 	}
 
-	result, err := client.GenerateText(ctx, messages, infrallm.GenerateOptions{
+	result, err := client.GenerateText(ctx, messages, llmservice.GenerateOptions{
 		MaxTokens: 4000,
 	})
 	if err != nil {
diff --git a/backend/newAgent/router/decision_parser.go b/backend/newAgent/router/decision_parser.go
index d990552..dd46671 100644
--- a/backend/newAgent/router/decision_parser.go
+++ b/backend/newAgent/router/decision_parser.go
@@ -26,7 +26,7 @@ var (
 // StreamDecisionResult 描述解析器的最终输出状态。
 type StreamDecisionResult struct {
 	// DecisionJSON 是标签内提取的完整 JSON 字符串。
-	// 调用方应使用 infrallm.ParseJSONObject[T] 将其解析为具体决策类型。
+	// 调用方应使用 llmservice.ParseJSONObject[T] 将其解析为具体决策类型。
 	DecisionJSON string
 
 	// BeforeText 是 <SMARTFLOW_DECISION> 标签之前的自然语言前言。
@@ -179,7 +179,7 @@ func (p *StreamDecisionParser) Result() *StreamDecisionResult {
 }
 
 // extractJSONFromTag 从标签内文本中提取第一个完整 JSON 对象。
-// 复用括号计数逻辑，与 infrallm.ExtractJSONObject 一致。
+// 复用括号计数逻辑，与 llmservice.ExtractJSONObject 一致。
 func extractJSONFromTag(text string) string {
 	clean := strings.TrimSpace(text)
 	if clean == "" {
diff --git a/backend/newAgent/shared/node_thinking.go b/backend/newAgent/shared/node_thinking.go
index ad14b7c..9a2da15 100644
--- a/backend/newAgent/shared/node_thinking.go
+++ b/backend/newAgent/shared/node_thinking.go
@@ -1,10 +1,10 @@
 package newagentshared
 
-import infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
+import llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 
-func ResolveThinkingMode(enabled bool) infrallm.ThinkingMode {
+func ResolveThinkingMode(enabled bool) llmservice.ThinkingMode {
 	if enabled {
-		return infrallm.ThinkingModeEnabled
+		return llmservice.ThinkingModeEnabled
 	}
-	return infrallm.ThinkingModeDisabled
+	return llmservice.ThinkingModeDisabled
 }
diff --git a/backend/newAgent/shared/node_unified_compact.go b/backend/newAgent/shared/node_unified_compact.go
index 7d1c41f..4dab8c6 100644
--- a/backend/newAgent/shared/node_unified_compact.go
+++ b/backend/newAgent/shared/node_unified_compact.go
@@ -6,11 +6,11 @@ import (
 	"fmt"
 	"log"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
 	newagentmodel "github.com/LoveLosita/smartflow/backend/newAgent/model"
 	newagentprompt "github.com/LoveLosita/smartflow/backend/newAgent/prompt"
 	newagentstream "github.com/LoveLosita/smartflow/backend/newAgent/stream"
 	"github.com/LoveLosita/smartflow/backend/pkg"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 	"github.com/cloudwego/eino/schema"
 )
 
@@ -22,7 +22,7 @@ import (
 // 3. StageName 和 StatusBlockID 用于区分日志来源与 SSE 状态推送目标。
 type UnifiedCompactInput struct {
 	// Client 用于调用 LLM 压缩 msg1/msg2。
-	Client *infrallm.Client
+	Client *llmservice.Client
 	// CompactionStore 用于持久化压缩摘要和 token 统计，为 nil 时跳过持久化。
 	CompactionStore newagentmodel.CompactionStore
 	// FlowState 提供 userID / conversationID / roundUsed 等定位信息。
diff --git a/backend/newAgent/stream/emitter.go b/backend/newAgent/stream/emitter.go
index b0fb041..6717557 100644
--- a/backend/newAgent/stream/emitter.go
+++ b/backend/newAgent/stream/emitter.go
@@ -8,7 +8,7 @@ import (
 	"sync"
 	"time"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 )
 
 // PayloadEmitter 是真正向外层 SSE 管道写 chunk 的最小接口。
@@ -540,7 +540,7 @@ func (e *ChunkEmitter) EmitDone() error {
 // 3. 不负责打开/关闭 StreamReader，调用方负责生命周期管理。
 func (e *ChunkEmitter) EmitStreamAssistantText(
 	ctx context.Context,
-	reader infrallm.StreamReader,
+	reader llmservice.StreamReader,
 	blockID, stage string,
 ) (string, error) {
 	if e == nil || reader == nil {
@@ -598,7 +598,7 @@ func (e *ChunkEmitter) EmitStreamAssistantText(
 // 用于只需展示思考过程而无需展示正文的场景。
 func (e *ChunkEmitter) EmitStreamReasoningText(
 	ctx context.Context,
-	reader infrallm.StreamReader,
+	reader llmservice.StreamReader,
 	blockID, stage string,
 ) (string, error) {
 	if e == nil || reader == nil {
diff --git a/backend/newAgent/tools/registry.go b/backend/newAgent/tools/registry.go
index 5c55e57..d91dfcd 100644
--- a/backend/newAgent/tools/registry.go
+++ b/backend/newAgent/tools/registry.go
@@ -5,9 +5,9 @@ import (
 	"sort"
 	"strings"
 
-	infrarag "github.com/LoveLosita/smartflow/backend/infra/rag"
 	"github.com/LoveLosita/smartflow/backend/newAgent/tools/schedule"
 	"github.com/LoveLosita/smartflow/backend/newAgent/tools/web"
+	ragservice "github.com/LoveLosita/smartflow/backend/services/rag"
 )
 
 // ToolHandler 约定所有工具的统一执行签名。
@@ -32,7 +32,7 @@ type ToolSchemaEntry struct {
 // 2. 某些依赖即便暂未使用也允许保留，避免业务层重新到处 new；
 // 3. 具体依赖缺失时由对应工具自行返回结构化失败结果。
 type DefaultRegistryDeps struct {
-	RAGRuntime infrarag.Runtime
+	RAGRuntime ragservice.Runtime
 
 	// WebSearchProvider 为 nil 时，web_search / web_fetch 仍会注册，
 	// 但 handler 会返回“暂未启用”的只读 observation，不阻断主流程。
diff --git a/backend/service/agent_bridge.go b/backend/service/agent_bridge.go
index 1f7b118..a436ee8 100644
--- a/backend/service/agent_bridge.go
+++ b/backend/service/agent_bridge.go
@@ -3,8 +3,8 @@ package service
 import (
 	"github.com/LoveLosita/smartflow/backend/dao"
 	outboxinfra "github.com/LoveLosita/smartflow/backend/infra/outbox"
-	"github.com/LoveLosita/smartflow/backend/inits"
 	"github.com/LoveLosita/smartflow/backend/service/agentsvc"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 )
 
 // AgentService 是 service 层对 agentsvc.AgentService 的兼容别名。
@@ -20,7 +20,7 @@ type AgentService = agentsvc.AgentService
 // 2) 主动调度 session DAO 也在这里显式透传，避免聊天入口再去回查全局单例；
 // 3) 真实构造逻辑已下沉到 service/agentsvc 包。
 func NewAgentService(
-	aiHub *inits.AIHub,
+	llmService *llmservice.Service,
 	repo *dao.AgentDAO,
 	taskRepo *dao.TaskDAO,
 	cacheDAO *dao.CacheDAO,
@@ -29,7 +29,7 @@ func NewAgentService(
 	activeSessionDAO *dao.ActiveScheduleSessionDAO,
 	eventPublisher outboxinfra.EventPublisher,
 ) *AgentService {
-	return agentsvc.NewAgentService(aiHub, repo, taskRepo, cacheDAO, agentRedis, activeScheduleDAO, activeSessionDAO, eventPublisher)
+	return agentsvc.NewAgentService(llmService, repo, taskRepo, cacheDAO, agentRedis, activeScheduleDAO, activeSessionDAO, eventPublisher)
 }
 
 // NewAgentServiceWithSchedule 在基础 AgentService 上注入排程依赖。
@@ -39,7 +39,7 @@ func NewAgentService(
 // 2) 排程依赖为可选：未注入时排程路由自动回退到普通聊天；
 // 3) 主动调度 session DAO 仍沿用统一构造注入，避免排程分支自己拼装仓储。
 func NewAgentServiceWithSchedule(
-	aiHub *inits.AIHub,
+	llmService *llmservice.Service,
 	repo *dao.AgentDAO,
 	taskRepo *dao.TaskDAO,
 	cacheDAO *dao.CacheDAO,
@@ -50,7 +50,7 @@ func NewAgentServiceWithSchedule(
 	scheduleSvc *ScheduleService,
 	taskSvc *TaskService,
 ) *AgentService {
-	svc := agentsvc.NewAgentService(aiHub, repo, taskRepo, cacheDAO, agentRedis, activeScheduleDAO, activeSessionDAO, eventPublisher)
+	svc := agentsvc.NewAgentService(llmService, repo, taskRepo, cacheDAO, agentRedis, activeScheduleDAO, activeSessionDAO, eventPublisher)
 
 	// 注入排程依赖：将 service 层方法包装为函数闭包，避免循环依赖。
 	if scheduleSvc != nil {
diff --git a/backend/service/agentsvc/agent.go b/backend/service/agentsvc/agent.go
index 9db635e..f50c2d4 100644
--- a/backend/service/agentsvc/agent.go
+++ b/backend/service/agentsvc/agent.go
@@ -3,6 +3,7 @@ package agentsvc
 import (
 	"context"
 	"encoding/json"
+	"errors"
 	"log"
 	"strconv"
 	"strings"
@@ -11,7 +12,6 @@ import (
 	"github.com/LoveLosita/smartflow/backend/conv"
 	"github.com/LoveLosita/smartflow/backend/dao"
 	outboxinfra "github.com/LoveLosita/smartflow/backend/infra/outbox"
-	"github.com/LoveLosita/smartflow/backend/inits"
 	memorymodel "github.com/LoveLosita/smartflow/backend/memory/model"
 	memoryobserve "github.com/LoveLosita/smartflow/backend/memory/observe"
 	"github.com/LoveLosita/smartflow/backend/model"
@@ -20,13 +20,13 @@ import (
 	newagenttools "github.com/LoveLosita/smartflow/backend/newAgent/tools"
 	"github.com/LoveLosita/smartflow/backend/pkg"
 	eventsvc "github.com/LoveLosita/smartflow/backend/service/events"
-	"github.com/cloudwego/eino-ext/components/model/ark"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 	"github.com/cloudwego/eino/schema"
 	"github.com/google/uuid"
 )
 
 type AgentService struct {
-	AIHub                    *inits.AIHub
+	llmService               *llmservice.Service
 	repo                     *dao.AgentDAO
 	taskRepo                 *dao.TaskDAO
 	cacheDAO                 *dao.CacheDAO
@@ -75,7 +75,7 @@ type AgentService struct {
 // 这里通过依赖注入把“模型、仓储、缓存、异步持久化通道”统一交给服务层管理，
 // 便于后续在单测中替换实现，或在启动流程中按环境切换配置。
 func NewAgentService(
-	aiHub *inits.AIHub,
+	llmService *llmservice.Service,
 	repo *dao.AgentDAO,
 	taskRepo *dao.TaskDAO,
 	cacheDAO *dao.CacheDAO,
@@ -90,7 +90,7 @@ func NewAgentService(
 	ensureTokenMeterCallbackRegistered()
 
 	return &AgentService{
-		AIHub:                    aiHub,
+		llmService:               llmService,
 		repo:                     repo,
 		taskRepo:                 taskRepo,
 		cacheDAO:                 cacheDAO,
@@ -123,8 +123,11 @@ func thinkingModeToBool(mode string) bool {
 // 当前约定：
 // - 旧链路已全面切到 newAgent graph，这里仅作为 runNormalChatFlow 回退时的模型选择入口；
 // - 统一返回 Pro 模型，旧 strategist 参数不再生效。
-func (s *AgentService) pickChatModel(requestModel string) (*ark.ChatModel, string) {
-	return s.AIHub.Pro, "pro"
+func (s *AgentService) pickChatModel(requestModel string) (*llmservice.Client, string) {
+	if s == nil || s.llmService == nil {
+		return nil, "pro"
+	}
+	return s.llmService.ProClient(), "pro"
 }
 
 // PersistChatHistory 是 Agent 聊天链路唯一的“消息持久化入口”。
@@ -304,7 +307,7 @@ func pushErrNonBlocking(errChan chan error, err error) {
 // 2) 开启随口记进度推送后，最终判定“非随口记”时回落到普通聊天。
 func (s *AgentService) runNormalChatFlow(
 	ctx context.Context,
-	selectedModel *ark.ChatModel,
+	selectedModel *llmservice.Client,
 	resolvedModelName string,
 	userMessage string,
 	userPersisted bool,
@@ -365,6 +368,12 @@ func (s *AgentService) runNormalChatFlow(
 		}
 	}
 
+	// 6.0. 没有可用模型时，直接中止普通聊天，避免写入半截用户消息后没有后续回复。
+	if selectedModel == nil {
+		pushErrNonBlocking(errChan, errors.New("llm client is not ready"))
+		return
+	}
+
 	// 6. 执行真正的流式聊天。
 	//    fullText 用于后续写 Redis/持久化，outChan 用于把流片段实时推给前端。
 	fullText, _, reasoningDurationSeconds, streamUsage, streamErr := s.streamChatFallback(ctx, selectedModel, resolvedModelName, userMessage, ifThinking, chatHistory, outChan, assistantReasoningStartedAt, userID, chatID)
diff --git a/backend/service/agentsvc/agent_meta.go b/backend/service/agentsvc/agent_meta.go
index c3ea28f..4122308 100644
--- a/backend/service/agentsvc/agent_meta.go
+++ b/backend/service/agentsvc/agent_meta.go
@@ -11,10 +11,8 @@ import (
 	"github.com/LoveLosita/smartflow/backend/model"
 	"github.com/LoveLosita/smartflow/backend/respond"
 	eventsvc "github.com/LoveLosita/smartflow/backend/service/events"
-	"github.com/cloudwego/eino-ext/components/model/ark"
-	einoModel "github.com/cloudwego/eino/components/model"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 	"github.com/cloudwego/eino/schema"
-	arkModel "github.com/volcengine/volcengine-go-sdk/service/arkruntime/model"
 )
 
 const (
@@ -253,11 +251,11 @@ func (s *AgentService) generateConversationTitle(ctx context.Context, history []
 	}
 
 	// 2. 标题生成属于结构化短输出，关闭 thinking 并限制 tokens，降低延迟与发散。
-	resp, err := modelInst.Generate(ctx, messages,
-		ark.WithThinking(&arkModel.Thinking{Type: arkModel.ThinkingTypeDisabled}),
-		einoModel.WithTemperature(0.2),
-		einoModel.WithMaxTokens(40),
-	)
+	resp, err := modelInst.GenerateText(ctx, messages, llmservice.GenerateOptions{
+		Temperature: 0.2,
+		MaxTokens:   40,
+		Thinking:    llmservice.ThinkingModeDisabled,
+	})
 	if err != nil {
 		return "", 0, err
 	}
@@ -267,26 +265,26 @@ func (s *AgentService) generateConversationTitle(ctx context.Context, history []
 
 	// 2.1 标题链路的 token 从模型响应 usage 中提取；缺失则按 0 处理，不影响主流程。
 	titleTokens := 0
-	if resp.ResponseMeta != nil && resp.ResponseMeta.Usage != nil {
+	if resp.Usage != nil {
 		titleTokens = normalizeUsageTotal(
-			resp.ResponseMeta.Usage.TotalTokens,
-			resp.ResponseMeta.Usage.PromptTokens,
-			resp.ResponseMeta.Usage.CompletionTokens,
+			resp.Usage.TotalTokens,
+			resp.Usage.PromptTokens,
+			resp.Usage.CompletionTokens,
 		)
 	}
-	return normalizeConversationTitle(resp.Content), titleTokens, nil
+	return normalizeConversationTitle(resp.Text), titleTokens, nil
 }
 
 // pickTitleModel 选择用于标题生成的模型。
 // 优先 Lite（成本低、速度快）；Lite 不可用时回退 Pro。
-func (s *AgentService) pickTitleModel() *ark.ChatModel {
-	if s.AIHub == nil {
+func (s *AgentService) pickTitleModel() *llmservice.Client {
+	if s == nil || s.llmService == nil {
 		return nil
 	}
-	if s.AIHub.Lite != nil {
-		return s.AIHub.Lite
+	if client := s.llmService.LiteClient(); client != nil {
+		return client
 	}
-	return s.AIHub.Pro
+	return s.llmService.ProClient()
 }
 
 // buildConversationTitleUserPrompt 把消息历史拼成可读文本供模型总结。
diff --git a/backend/service/agentsvc/agent_newagent.go b/backend/service/agentsvc/agent_newagent.go
index 0d56ab4..1a3bdb3 100644
--- a/backend/service/agentsvc/agent_newagent.go
+++ b/backend/service/agentsvc/agent_newagent.go
@@ -8,7 +8,6 @@ import (
 	"strings"
 	"time"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
 	newagentconv "github.com/LoveLosita/smartflow/backend/newAgent/conv"
 	newagentgraph "github.com/LoveLosita/smartflow/backend/newAgent/graph"
 	newagentmodel "github.com/LoveLosita/smartflow/backend/newAgent/model"
@@ -57,6 +56,11 @@ func (s *AgentService) runNewAgentGraph(
 	errChan chan error,
 ) {
 	requestCtx, _ := withRequestTokenMeter(ctx)
+	if s == nil || s.llmService == nil {
+		// 0. newAgent 主链强依赖 llm-service；装配漏传时直接返回错误，避免 nil receiver panic。
+		pushErrNonBlocking(errChan, errors.New("agent llm service is not initialized"))
+		return
+	}
 
 	// 1. 规范会话 ID 和模型选择。
 	chatID = normalizeConversationID(chatID)
@@ -184,14 +188,15 @@ func (s *AgentService) runNewAgentGraph(
 	}
 	graphRequest.Normalize()
 
-	// 8. 适配 LLM clients（从 AIHub 的 ark.ChatModel 转换为 newAgent LLM Client）。
+	// 8. 适配 LLM clients（统一从 llm-service 取出 newAgent 图所需模型，不再直接碰 AIHub）。
 	// 8.1 Chat/Deliver 使用 Pro 模型：路由分流、闲聊、交付总结属于标准复杂度。
 	// 8.2 Plan/Execute 使用 Max 模型：规划和 ReAct 循环需要深度推理能力。
-	chatClient := infrallm.WrapArkClient(s.AIHub.Pro)
-	planClient := infrallm.WrapArkClient(s.AIHub.Max)
-	executeClient := infrallm.WrapArkClient(s.AIHub.Max)
-	deliverClient := infrallm.WrapArkClient(s.AIHub.Pro)
-	summaryClient := infrallm.WrapArkClient(s.AIHub.Lite)
+	llmClients := s.llmService.NewAgentModelClients()
+	chatClient := llmClients.Chat
+	planClient := llmClients.Plan
+	executeClient := llmClients.Execute
+	deliverClient := llmClients.Deliver
+	summaryClient := llmClients.Summary
 
 	// 9. 适配 SSE emitter。
 	sseEmitter := newagentstream.NewSSEPayloadEmitter(outChan)
@@ -244,8 +249,8 @@ func (s *AgentService) runNewAgentGraph(
 		log.Printf("[ERROR] newAgent graph 执行失败 trace=%s chat=%s: %v", traceID, chatID, graphErr)
 		pushErrNonBlocking(errChan, fmt.Errorf("graph 执行失败: %w", graphErr))
 
-		// Graph 出错时回退普通聊天，保证可用性。回退使用 Pro 模型。
-		s.runNormalChatFlow(requestCtx, s.AIHub.Pro, resolvedModelName, userMessage, true, "", nil, thinkingModeToBool(thinkingMode), userID, chatID, traceID, requestStart, outChan, errChan)
+		// Graph 出错时回退普通聊天，保证可用性。回退使用 llm-service 的 Pro 模型。
+		s.runNormalChatFlow(requestCtx, chatClient, resolvedModelName, userMessage, true, "", nil, thinkingModeToBool(thinkingMode), userID, chatID, traceID, requestStart, outChan, errChan)
 		return
 	}
 
diff --git a/backend/service/agentsvc/agent_stream_fallback.go b/backend/service/agentsvc/agent_stream_fallback.go
index 38c5ef6..43be6f6 100644
--- a/backend/service/agentsvc/agent_stream_fallback.go
+++ b/backend/service/agentsvc/agent_stream_fallback.go
@@ -6,20 +6,18 @@ import (
 	"strings"
 	"time"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
 	newagentprompt "github.com/LoveLosita/smartflow/backend/newAgent/prompt"
 	newagentstream "github.com/LoveLosita/smartflow/backend/newAgent/stream"
-	"github.com/cloudwego/eino-ext/components/model/ark"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 	"github.com/cloudwego/eino/schema"
 	"github.com/google/uuid"
-	arkModel "github.com/volcengine/volcengine-go-sdk/service/arkruntime/model"
 )
 
 // streamChatFallback 是 graph 执行失败时的降级流式聊天。
 // 内联了旧 agentchat.StreamChat 的核心逻辑，不再依赖 agent/ 包。
 func (s *AgentService) streamChatFallback(
 	ctx context.Context,
-	llm *ark.ChatModel,
+	llm *llmservice.Client,
 	modelName string,
 	userInput string,
 	ifThinking bool,
@@ -36,13 +34,6 @@ func (s *AgentService) streamChatFallback(
 	}
 	messages = append(messages, schema.UserMessage(userInput))
 
-	var thinking *ark.Thinking
-	if ifThinking {
-		thinking = &arkModel.Thinking{Type: arkModel.ThinkingTypeEnabled}
-	} else {
-		thinking = &arkModel.Thinking{Type: arkModel.ThinkingTypeDisabled}
-	}
-
 	if strings.TrimSpace(modelName) == "" {
 		modelName = "smartflow-worker"
 	}
@@ -50,7 +41,11 @@ func (s *AgentService) streamChatFallback(
 	created := time.Now().Unix()
 	firstChunk := true
 	chunkEmitter := newagentstream.NewChunkEmitter(newagentstream.NewSSEPayloadEmitter(outChan), requestID, modelName, created)
-	chunkEmitter.SetReasoningSummaryFunc(s.makeReasoningSummaryFunc(infrallm.WrapArkClient(s.AIHub.Lite)))
+	reasoningSummaryClient := s.llmService.LiteClient()
+	if reasoningSummaryClient == nil {
+		reasoningSummaryClient = s.llmService.ProClient()
+	}
+	chunkEmitter.SetReasoningSummaryFunc(s.makeReasoningSummaryFunc(reasoningSummaryClient))
 	chunkEmitter.SetExtraEventHook(func(extra *newagentstream.OpenAIChunkExtra) {
 		s.persistNewAgentTimelineExtraEvent(context.Background(), userID, chatID, extra)
 	})
@@ -75,7 +70,14 @@ func (s *AgentService) streamChatFallback(
 	}
 	var reasoningEndAt *time.Time
 
-	reader, err := llm.Stream(ctx, messages, ark.WithThinking(thinking))
+	thinkingMode := llmservice.ThinkingModeDisabled
+	if ifThinking {
+		thinkingMode = llmservice.ThinkingModeEnabled
+	}
+
+	reader, err := llm.Stream(ctx, messages, llmservice.GenerateOptions{
+		Thinking: thinkingMode,
+	})
 	if err != nil {
 		return "", "", 0, nil, err
 	}
diff --git a/backend/service/agentsvc/reasoning_summary.go b/backend/service/agentsvc/reasoning_summary.go
index b2b1588..d75a7d3 100644
--- a/backend/service/agentsvc/reasoning_summary.go
+++ b/backend/service/agentsvc/reasoning_summary.go
@@ -6,9 +6,9 @@ import (
 	"log"
 	"strings"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
 	newagentprompt "github.com/LoveLosita/smartflow/backend/newAgent/prompt"
 	newagentstream "github.com/LoveLosita/smartflow/backend/newAgent/stream"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 )
 
 const reasoningSummaryMaxTokens = 700
@@ -24,7 +24,7 @@ type reasoningSummaryLLMResponse struct {
 // 1. service 层负责选择模型与 prompt，stream 层只负责调度和闸门；
 // 2. 这里不持久化摘要，持久化统一走 ChunkEmitter 的 extra hook；
 // 3. 摘要失败时返回 error，由 ReasoningDigestor 吞掉并等待下一次水位线/Flush 兜底。
-func (s *AgentService) makeReasoningSummaryFunc(client *infrallm.Client) newagentstream.ReasoningSummaryFunc {
+func (s *AgentService) makeReasoningSummaryFunc(client *llmservice.Client) newagentstream.ReasoningSummaryFunc {
 	if client == nil {
 		return nil
 	}
@@ -47,14 +47,14 @@ func (s *AgentService) makeReasoningSummaryFunc(client *infrallm.Client) newagen
 			DurationSeconds: input.DurationSeconds,
 		})
 
-		resp, rawResult, err := infrallm.GenerateJSON[reasoningSummaryLLMResponse](
+		resp, rawResult, err := llmservice.GenerateJSON[reasoningSummaryLLMResponse](
 			ctx,
 			client,
 			messages,
-			infrallm.GenerateOptions{
+			llmservice.GenerateOptions{
 				Temperature: 0.1,
 				MaxTokens:   reasoningSummaryMaxTokens,
-				Thinking:    infrallm.ThinkingModeDisabled,
+				Thinking:    llmservice.ThinkingModeDisabled,
 				Metadata: map[string]any{
 					"stage":         "reasoning_summary",
 					"candidate_seq": input.CandidateSeq,
@@ -99,7 +99,7 @@ func limitReasoningDetailSummary(text string, maxRunes int) string {
 	return string(runes[:maxRunes])
 }
 
-func truncateReasoningSummaryRaw(raw *infrallm.TextResult) string {
+func truncateReasoningSummaryRaw(raw *llmservice.TextResult) string {
 	if raw == nil {
 		return ""
 	}
diff --git a/backend/service/course.go b/backend/service/course.go
index d7cd299..9dd87d6 100644
--- a/backend/service/course.go
+++ b/backend/service/course.go
@@ -6,16 +6,16 @@ import (
 
 	"github.com/LoveLosita/smartflow/backend/conv"
 	"github.com/LoveLosita/smartflow/backend/dao"
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
 	"github.com/LoveLosita/smartflow/backend/model"
 	"github.com/LoveLosita/smartflow/backend/respond"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 )
 
 type CourseService struct {
 	// 伸出手：准备接住 DAO
 	courseDAO                  *dao.CourseDAO
 	scheduleDAO                *dao.ScheduleDAO
-	courseImageResponsesClient *infrallm.ArkResponsesClient
+	courseImageResponsesClient *llmservice.ArkResponsesClient
 	courseImageConfig          CourseImageParseConfig
 	courseImageModel           string
 }
@@ -24,7 +24,7 @@ type CourseService struct {
 func NewCourseService(
 	courseDAO *dao.CourseDAO,
 	scheduleDAO *dao.ScheduleDAO,
-	courseImageResponsesClient *infrallm.ArkResponsesClient,
+	courseImageResponsesClient *llmservice.ArkResponsesClient,
 	courseImageConfig CourseImageParseConfig,
 	courseImageModel string,
 ) *CourseService {
diff --git a/backend/service/course_parse_ark.go b/backend/service/course_parse_ark.go
index 5fc338f..45ea9c7 100644
--- a/backend/service/course_parse_ark.go
+++ b/backend/service/course_parse_ark.go
@@ -8,16 +8,20 @@ import (
 	"strings"
 	"time"
 
-	infrallm "github.com/LoveLosita/smartflow/backend/infra/llm"
 	"github.com/LoveLosita/smartflow/backend/model"
+	llmservice "github.com/LoveLosita/smartflow/backend/services/llm"
 )
 
 // ParseCourseTableImage 使用 Ark SDK Responses 解析课程表图片。
 func (ss *CourseService) ParseCourseTableImage(ctx context.Context, req model.CourseImageParseRequest) (*model.CourseImageParseResponse, error) {
 	if ss == nil || ss.courseImageResponsesClient == nil {
+		modelName := ""
+		if ss != nil {
+			modelName = ss.courseImageModel
+		}
 		log.Printf(
 			"[COURSE_PARSE][SERVICE] parser unavailable model_name=%q filename=%q mime=%q bytes=%d",
-			ss.courseImageModel,
+			modelName,
 			req.Filename,
 			req.MIMEType,
 			len(req.ImageBytes),
@@ -57,7 +61,7 @@ func (ss *CourseService) ParseCourseTableImage(ctx context.Context, req model.Co
 		base64Chars,
 		promptChars,
 		base64Chars+promptChars+len(strings.TrimSpace(courseImageParseSystemPrompt)),
-		infrallm.ThinkingModeDisabled,
+		llmservice.ThinkingModeDisabled,
 		courseImageParseTemperature,
 		ss.courseImageConfig.MaxTokens,
 		"json_object",
@@ -66,10 +70,10 @@ func (ss *CourseService) ParseCourseTableImage(ctx context.Context, req model.Co
 	// 1. 课程表图片识别输出体量大，显式透传 max_output_tokens，避免被默认值截断。
 	// 2. text_format 固定为 json_object，降低输出混入解释文本导致解析失败的概率。
 	// 3. thinking 显式关闭，优先保证课程导入链路稳定性。
-	draft, rawResult, err := infrallm.GenerateArkResponsesJSON[model.CourseImageParseResponse](ctx, ss.courseImageResponsesClient, messages, infrallm.ArkResponsesOptions{
+	draft, rawResult, err := llmservice.GenerateArkResponsesJSON[model.CourseImageParseResponse](ctx, ss.courseImageResponsesClient, messages, llmservice.ArkResponsesOptions{
 		Temperature:     courseImageParseTemperature,
 		MaxOutputTokens: ss.courseImageConfig.MaxTokens,
-		Thinking:        infrallm.ThinkingModeDisabled,
+		Thinking:        llmservice.ThinkingModeDisabled,
 		TextFormat:      "json_object",
 	})
 	if err != nil {
@@ -188,12 +192,12 @@ func (ss *CourseService) ParseCourseTableImage(ctx context.Context, req model.Co
 	return normalizedDraft, nil
 }
 
-func buildCourseImageParseResponsesMessages(req *model.CourseImageParseRequest) ([]infrallm.ArkResponsesMessage, int, int) {
+func buildCourseImageParseResponsesMessages(req *model.CourseImageParseRequest) ([]llmservice.ArkResponsesMessage, int, int) {
 	userPrompt := fmt.Sprintf(courseImageParseUserPromptTemplate, req.Filename, req.MIMEType)
 	base64Data := base64.StdEncoding.EncodeToString(req.ImageBytes)
 	imageDataURL := fmt.Sprintf("data:%s;base64,%s", req.MIMEType, base64Data)
 
-	messages := []infrallm.ArkResponsesMessage{
+	messages := []llmservice.ArkResponsesMessage{
 		{
 			Role: "system",
 			Text: strings.TrimSpace(courseImageParseSystemPrompt),
@@ -208,7 +212,7 @@ func buildCourseImageParseResponsesMessages(req *model.CourseImageParseRequest)
 	return messages, len(base64Data), len(strings.TrimSpace(userPrompt))
 }
 
-func isCourseImageOutputTruncated(rawResult *infrallm.ArkResponsesResult) bool {
+func isCourseImageOutputTruncated(rawResult *llmservice.ArkResponsesResult) bool {
 	if rawResult == nil {
 		return false
 	}
diff --git a/backend/infra/llm/ark.go b/backend/services/llm/ark.go
similarity index 73%
rename from backend/infra/llm/ark.go
rename to backend/services/llm/ark.go
index 6ee6c14..a2e7bbb 100644
--- a/backend/infra/llm/ark.go
+++ b/backend/services/llm/ark.go
@@ -1,7 +1,3 @@
-// 过渡期统一 Ark 调用封装。
-//
-// 这里保留 CallArkText / CallArkJSON，方便暂时还直接持有 *ark.ChatModel 的调用点
-// 逐步迁移到统一 Client。后续 memory 也可以直接复用这套中立层。
 package llm
 
 import (
@@ -15,12 +11,7 @@ import (
 	arkModel "github.com/volcengine/volcengine-go-sdk/service/arkruntime/model"
 )
 
-// ArkCallOptions 是基于 ark.ChatModel 的通用调用选项。
-//
-// 设计目的：
-// 1. 先把 Ark 调用样板抽成公共层；
-// 2. 再由 WrapArkClient 提供统一 Client；
-// 3. 让上层尽量只关注业务 prompt 和结构化结果。
+// ArkCallOptions 是直接调用 ark.ChatModel 时使用的通用入参。
 type ArkCallOptions struct {
 	Temperature float64
 	MaxTokens   int
@@ -28,12 +19,6 @@ type ArkCallOptions struct {
 }
 
 // CallArkText 调用 ark 模型并返回纯文本。
-//
-// 职责边界：
-// 1. 负责拼 system + user 两段消息；
-// 2. 负责统一配置 thinking / temperature / maxTokens；
-// 3. 负责拦截空响应；
-// 4. 不负责 JSON 解析，不负责业务字段校验。
 func CallArkText(ctx context.Context, chatModel *ark.ChatModel, systemPrompt, userPrompt string, options ArkCallOptions) (string, error) {
 	if chatModel == nil {
 		return "", errors.New("ark model is nil")
@@ -76,6 +61,7 @@ func buildArkOptions(options ArkCallOptions) []einoModel.Option {
 	if options.Thinking == ThinkingModeEnabled {
 		thinkingType = arkModel.ThinkingTypeEnabled
 	}
+
 	opts := []einoModel.Option{
 		ark.WithThinking(&arkModel.Thinking{Type: thinkingType}),
 		einoModel.WithTemperature(float32(options.Temperature)),
diff --git a/backend/infra/llm/ark_adapter.go b/backend/services/llm/ark_adapter.go
similarity index 68%
rename from backend/infra/llm/ark_adapter.go
rename to backend/services/llm/ark_adapter.go
index 3234501..9d2d28b 100644
--- a/backend/infra/llm/ark_adapter.go
+++ b/backend/services/llm/ark_adapter.go
@@ -12,17 +12,14 @@ import (
 )
 
 // WrapArkClient 将 ark.ChatModel 适配为统一 Client。
-//
-// 职责边界：
-// 1. generateText：调用 ark.ChatModel.Generate（非流式），供 GenerateJSON 使用；
-// 2. streamText：调用 ark.ChatModel.Stream（流式），供需要流式输出的场景使用；
-// 3. 两者共用同一套 options 转换。
+// 1. generateText 走 Generate，供 GenerateJSON/GenerateText 使用。
+// 2. streamText 走 Stream，供需要流式输出的场景使用。
+// 3. 两条路径共用同一套参数转换逻辑。
 func WrapArkClient(arkChatModel *ark.ChatModel) *Client {
 	if arkChatModel == nil {
 		return nil
 	}
 
-	// 非流式文本生成，供 GenerateJSON / GenerateText 调用路径使用。
 	generateFunc := func(ctx context.Context, messages []*schema.Message, options GenerateOptions) (*TextResult, error) {
 		arkOpts := buildArkStreamOptions(options)
 		msg, err := arkChatModel.Generate(ctx, messages, arkOpts...)
@@ -47,7 +44,6 @@ func WrapArkClient(arkChatModel *ark.ChatModel) *Client {
 		}, nil
 	}
 
-	// 流式文本生成。
 	streamFunc := func(ctx context.Context, messages []*schema.Message, options GenerateOptions) (StreamReader, error) {
 		arkOpts := buildArkStreamOptions(options)
 		reader, err := arkChatModel.Stream(ctx, messages, arkOpts...)
@@ -60,11 +56,10 @@ func WrapArkClient(arkChatModel *ark.ChatModel) *Client {
 	return NewClient(generateFunc, streamFunc)
 }
 
-// buildArkStreamOptions 将统一 GenerateOptions 转换为 ark 的流式调用选项。
+// buildArkStreamOptions 将统一的 GenerateOptions 转换为 ark 的流式调用参数。
 func buildArkStreamOptions(options GenerateOptions) []einoModel.Option {
 	thinkingEnabled := options.Thinking == ThinkingModeEnabled
 
-	// Thinking
 	thinkingType := arkModel.ThinkingTypeDisabled
 	if thinkingEnabled {
 		thinkingType = arkModel.ThinkingTypeEnabled
@@ -73,16 +68,12 @@ func buildArkStreamOptions(options GenerateOptions) []einoModel.Option {
 		ark.WithThinking(&arkModel.Thinking{Type: thinkingType}),
 	}
 
-	// Temperature：thinking 模型强制要求 temperature=1，否则 API 静默忽略 thinking。
 	if thinkingEnabled {
 		opts = append(opts, einoModel.WithTemperature(1.0))
 	} else if options.Temperature > 0 {
 		opts = append(opts, einoModel.WithTemperature(float32(options.Temperature)))
 	}
 
-	// MaxTokens：thinking 模式下 thinking token 占用 max_tokens 预算，
-	// 调用方设定的值仅代表"期望输出长度"，实际预算需留出思考空间。
-	// 最低保障 16000，避免思考链被截断导致输出为空或非 JSON。
 	maxTokens := options.MaxTokens
 	if thinkingEnabled {
 		const minThinkingBudget = 16000
@@ -97,14 +88,12 @@ func buildArkStreamOptions(options GenerateOptions) []einoModel.Option {
 	return opts
 }
 
-// arkStreamReaderAdapter 适配 ark.ChatModel.Stream 返回的 reader。
-// ark.Stream 返回 schema.StreamReader[*schema.Message]，其 Close() 方法无返回值
-// 而我们的 StreamReader 接口要求 Close() error
+// arkStreamReaderAdapter 把 ark 的流式 reader 转成统一的 StreamReader 接口。
 type arkStreamReaderAdapter struct {
 	reader *schema.StreamReader[*schema.Message]
 }
 
-// Recv 转发到 ark reader 的 Recv 方法。
+// Recv 转发到底层 reader。
 func (r *arkStreamReaderAdapter) Recv() (*schema.Message, error) {
 	if r == nil || r.reader == nil {
 		return nil, io.EOF
@@ -112,8 +101,7 @@ func (r *arkStreamReaderAdapter) Recv() (*schema.Message, error) {
 	return r.reader.Recv()
 }
 
-// Close 转发到 ark reader 的 Close 方法。
-// ark 的 Close() 无返回值，我们适配为返回 nil
+// Close 适配 ark reader 的 Close 行为。
 func (r *arkStreamReaderAdapter) Close() error {
 	if r == nil || r.reader == nil {
 		return nil
diff --git a/backend/infra/llm/ark_responses_client.go b/backend/services/llm/ark_responses_client.go
similarity index 93%
rename from backend/infra/llm/ark_responses_client.go
rename to backend/services/llm/ark_responses_client.go
index a82aed6..7f7eede 100644
--- a/backend/infra/llm/ark_responses_client.go
+++ b/backend/services/llm/ark_responses_client.go
@@ -11,11 +11,6 @@ import (
 )
 
 // ArkResponsesMessage 描述一次 Responses 输入消息。
-//
-// 职责边界：
-// 1. 负责表达角色与多模态内容（文本/图片）；
-// 2. 不负责业务 prompt 生成；
-// 3. 不负责输出 JSON 的字段校验。
 type ArkResponsesMessage struct {
 	Role        string
 	Text        string
@@ -23,7 +18,7 @@ type ArkResponsesMessage struct {
 	ImageDetail string
 }
 
-// ArkResponsesOptions 描述 Responses 生成选项。
+// ArkResponsesOptions 描述 Responses 调用参数。
 type ArkResponsesOptions struct {
 	Model           string
 	Temperature     float64
@@ -32,14 +27,14 @@ type ArkResponsesOptions struct {
 	TextFormat      string
 }
 
-// ArkResponsesUsage 统一透传 token 使用量。
+// ArkResponsesUsage 统一转写 token usage。
 type ArkResponsesUsage struct {
 	InputTokens  int64
 	OutputTokens int64
 	TotalTokens  int64
 }
 
-// ArkResponsesResult 是 Ark Responses 的统一输出结构。
+// ArkResponsesResult 是 Responses 调用的统一输出结构。
 type ArkResponsesResult struct {
 	Text             string
 	Status           string
@@ -56,11 +51,9 @@ type ArkResponsesClient struct {
 }
 
 // NewArkResponsesClient 创建 Ark SDK Responses 客户端。
-//
-// 说明：
-// 1. model 为空时返回 nil，表示当前能力未启用；
-// 2. baseURL 为空时使用 SDK 默认地址；
-// 3. 仅负责客户端创建，不做连通性探测。
+// 1. model 为空时直接返回 nil，表示这条能力没有启用。
+// 2. baseURL 为空时使用 SDK 默认地址。
+// 3. 这里只负责本地构造，不做连通性探测。
 func NewArkResponsesClient(apiKey string, baseURL string, model string) *ArkResponsesClient {
 	model = strings.TrimSpace(model)
 	if model == "" {
@@ -104,7 +97,7 @@ func (c *ArkResponsesClient) GenerateText(ctx context.Context, messages []ArkRes
 	return result, nil
 }
 
-// GenerateArkResponsesJSON 先调用 Responses，再解析为 JSON 结构体。
+// GenerateArkResponsesJSON 先调用 Responses，再解析成 JSON 结构体。
 func GenerateArkResponsesJSON[T any](ctx context.Context, client *ArkResponsesClient, messages []ArkResponsesMessage, options ArkResponsesOptions) (*T, *ArkResponsesResult, error) {
 	if client == nil {
 		return nil, nil, errors.New("ark responses client is not ready")
diff --git a/backend/infra/llm/client.go b/backend/services/llm/client.go
similarity index 56%
rename from backend/infra/llm/client.go
rename to backend/services/llm/client.go
index 4536971..2abedac 100644
--- a/backend/infra/llm/client.go
+++ b/backend/services/llm/client.go
@@ -9,12 +9,7 @@ import (
 	"github.com/cloudwego/eino/schema"
 )
 
-// ThinkingMode 描述本次模型调用对 thinking 的期望。
-//
-// 职责边界：
-// 1. 这里只表达“调用方希望怎样配置推理模式”；
-// 2. 不直接绑定某个具体模型厂商的参数枚举；
-// 3. 真正如何把它翻译成 ark / OpenAI / 其他 provider 的 option，由后续适配层负责。
+// ThinkingMode 描述这次模型调用对 thinking 的期望。
 type ThinkingMode string
 
 const (
@@ -23,12 +18,7 @@ const (
 	ThinkingModeDisabled ThinkingMode = "disabled"
 )
 
-// GenerateOptions 是统一模型调用选项。
-//
-// 设计目的：
-// 1. 先把“每个 skill / worker 都会反复传的参数”收敛成一份结构；
-// 2. 让上层以后只表达“我要什么”，不再自己重复组织 option；
-// 3. 暂时不追求覆盖所有 provider 参数，先把最常用的几个公共位抽出来。
+// GenerateOptions 统一收敛文本调用时最常见的公共参数。
 type GenerateOptions struct {
 	Temperature float64
 	MaxTokens   int
@@ -36,40 +26,32 @@ type GenerateOptions struct {
 	Metadata    map[string]any
 }
 
-// TextResult 是统一文本生成结果。
-//
-// 职责边界：
-// 1. Text 保存模型最终返回的纯文本；
-// 2. Usage 保存本次调用的 token 使用量，供后续统一统计；
-// 3. 不负责 JSON 解析，不负责业务字段映射。
+// TextResult 保存一次文本生成的最终结果和 usage。
+// 1. Text 存放模型返回的纯文本。
+// 2. Usage 方便上层做统一统计。
+// 3. 这里不负责 JSON 解析，也不负责业务字段映射。
 type TextResult struct {
-	Text  string
-	Usage *schema.TokenUsage
-	// FinishReason 透传 provider 的停止原因，便于上层判断是否因 length 等原因被截断。
+	Text         string
+	Usage        *schema.TokenUsage
 	FinishReason string
 }
 
-// StreamReader 抽象了“可逐块 Recv 的流式返回器”。
-//
-// 之所以不直接依赖某个具体 SDK 的 reader 类型，是因为现在还处在骨架收敛阶段，
-// 后续接 ark、OpenAI 兼容层还是别的 provider，都可以往这个最小接口上适配。
+// StreamReader 抽象可以逐块读取消息的流式返回器。
 type StreamReader interface {
 	Recv() (*schema.Message, error)
 	Close() error
 }
 
-// TextGenerateFunc 是文本生成的统一适配函数签名。
+// TextGenerateFunc 定义统一文本生成函数签名。
 type TextGenerateFunc func(ctx context.Context, messages []*schema.Message, options GenerateOptions) (*TextResult, error)
 
-// StreamGenerateFunc 是流式生成的统一适配函数签名。
+// StreamGenerateFunc 定义统一流式生成函数签名。
 type StreamGenerateFunc func(ctx context.Context, messages []*schema.Message, options GenerateOptions) (StreamReader, error)
 
 // Client 是统一模型客户端门面。
-//
-// 职责边界：
-// 1. 负责把调用方的“模型调用意图”收敛到统一入口；
-// 2. 负责统一参数校验、空响应防御、GenerateJSON 复用；
-// 3. 不负责写 prompt，不负责业务 fallback，也不直接持有具体厂商 SDK 细节。
+// 1. 只做最小输入校验和空响应防御。
+// 2. 不负责 prompt 拼装，也不负责业务 fallback。
+// 3. 具体 provider 的细节由上层适配器收敛进来。
 type Client struct {
 	generateText TextGenerateFunc
 	streamText   StreamGenerateFunc
@@ -84,11 +66,6 @@ func NewClient(generateText TextGenerateFunc, streamText StreamGenerateFunc) *Cl
 }
 
 // GenerateText 执行一次统一文本生成。
-//
-// 职责边界：
-// 1. 负责做最小必要的入参校验；
-// 2. 负责统一拦截“模型空响应”这类公共问题；
-// 3. 不负责业务 prompt 拼接，也不负责把文本再映射成业务结构。
 func (c *Client) GenerateText(ctx context.Context, messages []*schema.Message, options GenerateOptions) (*TextResult, error) {
 	if c == nil || c.generateText == nil {
 		return nil, errors.New("llm client is not ready")
@@ -111,11 +88,6 @@ func (c *Client) GenerateText(ctx context.Context, messages []*schema.Message, o
 }
 
 // GenerateJSON 先走统一文本生成，再走统一 JSON 解析。
-//
-// 设计说明：
-// 1. 把“Generate -> 提取 JSON -> 反序列化”这段公共链路收敛起来；
-// 2. 上层只关心业务结构，不需要重复实现解析样板；
-// 3. 返回 parsed + rawResult，方便打点与回退时保留原文。
 func GenerateJSON[T any](ctx context.Context, client *Client, messages []*schema.Message, options GenerateOptions) (*T, *TextResult, error) {
 	result, err := client.GenerateText(ctx, messages, options)
 	if err != nil {
@@ -130,11 +102,6 @@ func GenerateJSON[T any](ctx context.Context, client *Client, messages []*schema
 }
 
 // Stream 打开统一流式调用入口。
-//
-// 职责边界：
-// 1. 只负责把“流式生成能力”暴露给上层；
-// 2. 不负责 chunk 到 OpenAI 协议的转换，那部分应放在 stream/；
-// 3. 不负责累计全文，也不负责 token 统计落库。
 func (c *Client) Stream(ctx context.Context, messages []*schema.Message, options GenerateOptions) (StreamReader, error) {
 	if c == nil || c.streamText == nil {
 		return nil, errors.New("llm stream client is not ready")
@@ -145,12 +112,7 @@ func (c *Client) Stream(ctx context.Context, messages []*schema.Message, options
 	return c.streamText(ctx, messages, options)
 }
 
-// BuildSystemUserMessages 构造最常见的“system + history + user”消息列表。
-//
-// 设计说明：
-// 1. 先把最稳定的消息编排方式沉淀下来，减少各业务域样板代码；
-// 2. 只做消息切片装配，不做 prompt 生成；
-// 3. 供 agent / memory 等多个能力域复用。
+// BuildSystemUserMessages 构造最常见的 system + history + user 消息列表。
 func BuildSystemUserMessages(systemPrompt string, history []*schema.Message, userPrompt string) []*schema.Message {
 	messages := make([]*schema.Message, 0, len(history)+2)
 	if strings.TrimSpace(systemPrompt) != "" {
@@ -165,7 +127,7 @@ func BuildSystemUserMessages(systemPrompt string, history []*schema.Message, use
 	return messages
 }
 
-// CloneUsage 深拷贝 token usage，避免后续多处累加时共享同一指针。
+// CloneUsage 深拷贝 token usage，避免后续累加时共享同一个指针。
 func CloneUsage(usage *schema.TokenUsage) *schema.TokenUsage {
 	if usage == nil {
 		return nil
@@ -174,12 +136,7 @@ func CloneUsage(usage *schema.TokenUsage) *schema.TokenUsage {
 	return &copied
 }
 
-// MergeUsage 合并两段 usage。
-//
-// 合并策略：
-// 1. 对“同一次调用不同流分片”的场景，取更大值作为最终值；
-// 2. 对“多次独立调用累计”的场景，应由上层显式做加法，而不是用这个函数；
-// 3. 该函数只适用于“同一次调用的分块 usage 收敛”。
+// MergeUsage 合并两段 usage，取各字段更大的值作为累计结果。
 func MergeUsage(base *schema.TokenUsage, incoming *schema.TokenUsage) *schema.TokenUsage {
 	if incoming == nil {
 		return CloneUsage(base)
@@ -207,7 +164,7 @@ func MergeUsage(base *schema.TokenUsage, incoming *schema.TokenUsage) *schema.To
 	return &merged
 }
 
-// FormatEmptyResponseError 统一生成“模型返回空结果”的错误文案。
+// FormatEmptyResponseError 统一模型空结果的错误文案。
 func FormatEmptyResponseError(scene string) error {
 	scene = strings.TrimSpace(scene)
 	if scene == "" {
diff --git a/backend/infra/llm/json.go b/backend/services/llm/json.go
similarity index 67%
rename from backend/infra/llm/json.go
rename to backend/services/llm/json.go
index 2d029ed..c402f6c 100644
--- a/backend/infra/llm/json.go
+++ b/backend/services/llm/json.go
@@ -7,12 +7,10 @@ import (
 	"strings"
 )
 
-// ParseJSONObject 解析模型返回中的 JSON 对象。
-//
-// 职责边界：
-// 1. 负责处理“模型输出前后夹杂解释文字 / markdown 代码块”的常见情况；
-// 2. 负责提取最外层 JSON object 并反序列化为目标结构；
-// 3. 不负责业务字段合法性校验，应由上层调用方自行校验。
+// ParseJSONObject 解析模型返回内容中的 JSON 对象。
+// 1. 先剥离常见的 markdown 代码块包装。
+// 2. 再从混合文本里提取最外层 JSON 对象。
+// 3. 这里只负责结构解析，不负责字段合法性校验。
 func ParseJSONObject[T any](raw string) (*T, error) {
 	clean := strings.TrimSpace(raw)
 	if clean == "" {
@@ -31,12 +29,7 @@ func ParseJSONObject[T any](raw string) (*T, error) {
 	return &out, nil
 }
 
-// ExtractJSONObject 从混合文本里提取第一个完整 JSON 对象。
-//
-// 设计说明：
-// 1. LLM 很容易输出“这里是结果：{...}”这种半结构化文本；
-// 2. 这里用括号计数而不是正则，避免嵌套对象一多就误截断；
-// 3. 目前只提取 object，不提取 array，因为当前契约基本都是对象。
+// ExtractJSONObject 从混合文本中提取第一个完整的 JSON 对象。
 func ExtractJSONObject(text string) string {
 	clean := trimMarkdownCodeFence(strings.TrimSpace(text))
 	if clean == "" {
@@ -94,9 +87,6 @@ func trimMarkdownCodeFence(text string) string {
 		return trimmed
 	}
 
-	// 1. 去掉首行 ```json / ```；
-	// 2. 若末行是 ```，一并去掉；
-	// 3. 中间正文保持原样，避免破坏 JSON 的换行结构。
 	body := lines[1:]
 	if len(body) > 0 && strings.TrimSpace(body[len(body)-1]) == "```" {
 		body = body[:len(body)-1]
diff --git a/backend/services/llm/service.go b/backend/services/llm/service.go
new file mode 100644
index 0000000..8936c06
--- /dev/null
+++ b/backend/services/llm/service.go
@@ -0,0 +1,109 @@
+package llm
+
+import (
+	"strings"
+
+	"github.com/LoveLosita/smartflow/backend/inits"
+)
+
+// Service 只负责统一暴露已经构造好的模型客户端，不负责 prompt 和业务编排。
+type Service struct {
+	liteClient                 *Client
+	proClient                  *Client
+	maxClient                  *Client
+	courseImageResponsesClient *ArkResponsesClient
+}
+
+// Options 描述 llm-service 初始化时需要接管的启动期依赖。
+// 1. AIHub 仍然是当前进程内 Ark ChatModel 的来源，但服务层只保存统一 Client。
+// 2. CourseImageResponsesClient 允许外部预先注入，便于测试或特殊启动路径复用。
+// 3. 某个字段为空时不报错，直接保留 nil，交给上层继续走兼容降级。
+type Options struct {
+	AIHub                      *inits.AIHub
+	APIKey                     string
+	BaseURL                    string
+	CourseVisionModel          string
+	CourseImageResponsesClient *ArkResponsesClient
+}
+
+// AgentModelClients 一次性暴露 newAgent 图常用的模型分配结果。
+type AgentModelClients struct {
+	Chat    *Client
+	Plan    *Client
+	Execute *Client
+	Deliver *Client
+	Summary *Client
+}
+
+// New 构造 llm-service。
+// 1. 不返回 error，是为了让上层继续按 nil 客户端做逐步降级。
+// 2. 只要 AIHub 已初始化，就把其中的 ChatModel 收敛成统一 Client。
+// 3. 课程图片解析客户端在这里统一构建，避免业务层直接依赖 Responses SDK。
+func New(opts Options) *Service {
+	svc := &Service{}
+
+	if opts.AIHub != nil {
+		svc.liteClient = WrapArkClient(opts.AIHub.Lite)
+		svc.proClient = WrapArkClient(opts.AIHub.Pro)
+		svc.maxClient = WrapArkClient(opts.AIHub.Max)
+	}
+
+	if opts.CourseImageResponsesClient != nil {
+		svc.courseImageResponsesClient = opts.CourseImageResponsesClient
+	} else {
+		apiKey := strings.TrimSpace(opts.APIKey)
+		baseURL := strings.TrimSpace(opts.BaseURL)
+		model := strings.TrimSpace(opts.CourseVisionModel)
+		if apiKey != "" && model != "" {
+			svc.courseImageResponsesClient = NewArkResponsesClient(apiKey, baseURL, model)
+		}
+	}
+
+	return svc
+}
+
+// LiteClient 返回低成本短输出模型客户端。
+func (s *Service) LiteClient() *Client {
+	if s == nil {
+		return nil
+	}
+	return s.liteClient
+}
+
+// ProClient 返回默认复杂对话模型客户端。
+func (s *Service) ProClient() *Client {
+	if s == nil {
+		return nil
+	}
+	return s.proClient
+}
+
+// MaxClient 返回深度推理模型客户端。
+func (s *Service) MaxClient() *Client {
+	if s == nil {
+		return nil
+	}
+	return s.maxClient
+}
+
+// CourseImageResponsesClient 返回课程图片解析所用的 Responses 客户端。
+func (s *Service) CourseImageResponsesClient() *ArkResponsesClient {
+	if s == nil {
+		return nil
+	}
+	return s.courseImageResponsesClient
+}
+
+// NewAgentModelClients 一次性返回 newAgent 图里常用的模型分配。
+func (s *Service) NewAgentModelClients() AgentModelClients {
+	if s == nil {
+		return AgentModelClients{}
+	}
+	return AgentModelClients{
+		Chat:    s.ProClient(),
+		Plan:    s.MaxClient(),
+		Execute: s.MaxClient(),
+		Deliver: s.ProClient(),
+		Summary: s.LiteClient(),
+	}
+}
diff --git a/backend/infra/rag/service.go b/backend/services/rag/api.go
similarity index 97%
rename from backend/infra/rag/service.go
rename to backend/services/rag/api.go
index f8bce78..1cdfca9 100644
--- a/backend/infra/rag/service.go
+++ b/backend/services/rag/api.go
@@ -5,7 +5,7 @@ import (
 	"time"
 )
 
-// Runtime 是 RAG Infra 对业务侧暴露的唯一稳定方法面。
+// Runtime 是 RAG service 对业务侧暴露的唯一稳定方法面。
 //
 // 职责边界：
 // 1. 负责承接 memory/web 两类语料的统一入库与检索入口；
diff --git a/backend/infra/rag/chunk/text_chunker.go b/backend/services/rag/chunk/text_chunker.go
similarity index 96%
rename from backend/infra/rag/chunk/text_chunker.go
rename to backend/services/rag/chunk/text_chunker.go
index 39ed133..fb61c5d 100644
--- a/backend/infra/rag/chunk/text_chunker.go
+++ b/backend/services/rag/chunk/text_chunker.go
@@ -5,7 +5,7 @@ import (
 	"fmt"
 	"strings"
 
-	"github.com/LoveLosita/smartflow/backend/infra/rag/core"
+	"github.com/LoveLosita/smartflow/backend/services/rag/core"
 )
 
 // TextChunker 是默认文本切块器。
diff --git a/backend/infra/rag/config/config.go b/backend/services/rag/config/config.go
similarity index 100%
rename from backend/infra/rag/config/config.go
rename to backend/services/rag/config/config.go
diff --git a/backend/infra/rag/core/errors.go b/backend/services/rag/core/errors.go
similarity index 100%
rename from backend/infra/rag/core/errors.go
rename to backend/services/rag/core/errors.go
diff --git a/backend/infra/rag/core/interfaces.go b/backend/services/rag/core/interfaces.go
similarity index 100%
rename from backend/infra/rag/core/interfaces.go
rename to backend/services/rag/core/interfaces.go
diff --git a/backend/infra/rag/core/observer.go b/backend/services/rag/core/observer.go
similarity index 97%
rename from backend/infra/rag/core/observer.go
rename to backend/services/rag/core/observer.go
index d96d4fd..c00c31e 100644
--- a/backend/infra/rag/core/observer.go
+++ b/backend/services/rag/core/observer.go
@@ -21,7 +21,7 @@ const (
 // ObserveEvent 描述一次统一观测事件。
 //
 // 职责边界：
-// 1. 只承载 RAG Infra 的结构化运行信息；
+// 1. 只承载 RAG service 的结构化运行信息；
 // 2. 不绑定具体日志系统、指标系统或 tracing 实现；
 // 3. 字段内容应尽量稳定，便于后续统一接入全局观测平台。
 type ObserveEvent struct {
@@ -31,7 +31,7 @@ type ObserveEvent struct {
 	Fields    map[string]any
 }
 
-// Observer 是 RAG Infra 的最小观测接口。
+// Observer 是 RAG service 的最小观测接口。
 //
 // 职责边界：
 // 1. 负责消费结构化事件；
diff --git a/backend/infra/rag/core/pipeline.go b/backend/services/rag/core/pipeline.go
similarity index 100%
rename from backend/infra/rag/core/pipeline.go
rename to backend/services/rag/core/pipeline.go
diff --git a/backend/infra/rag/core/types.go b/backend/services/rag/core/types.go
similarity index 100%
rename from backend/infra/rag/core/types.go
rename to backend/services/rag/core/types.go
diff --git a/backend/infra/rag/corpus/common.go b/backend/services/rag/corpus/common.go
similarity index 100%
rename from backend/infra/rag/corpus/common.go
rename to backend/services/rag/corpus/common.go
diff --git a/backend/infra/rag/corpus/memory_corpus.go b/backend/services/rag/corpus/memory_corpus.go
similarity index 98%
rename from backend/infra/rag/corpus/memory_corpus.go
rename to backend/services/rag/corpus/memory_corpus.go
index af461f8..62d9991 100644
--- a/backend/infra/rag/corpus/memory_corpus.go
+++ b/backend/services/rag/corpus/memory_corpus.go
@@ -7,7 +7,7 @@ import (
 	"strings"
 	"time"
 
-	"github.com/LoveLosita/smartflow/backend/infra/rag/core"
+	"github.com/LoveLosita/smartflow/backend/services/rag/core"
 )
 
 const memoryCorpusName = "memory"
diff --git a/backend/infra/rag/corpus/web_corpus.go b/backend/services/rag/corpus/web_corpus.go
similarity index 98%
rename from backend/infra/rag/corpus/web_corpus.go
rename to backend/services/rag/corpus/web_corpus.go
index a3f4443..fd90724 100644
--- a/backend/infra/rag/corpus/web_corpus.go
+++ b/backend/services/rag/corpus/web_corpus.go
@@ -7,7 +7,7 @@ import (
 	"strings"
 	"time"
 
-	"github.com/LoveLosita/smartflow/backend/infra/rag/core"
+	"github.com/LoveLosita/smartflow/backend/services/rag/core"
 )
 
 const webCorpusName = "web"
diff --git a/backend/infra/rag/embed/eino_embedder.go b/backend/services/rag/embed/eino_embedder.go
similarity index 100%
rename from backend/infra/rag/embed/eino_embedder.go
rename to backend/services/rag/embed/eino_embedder.go
diff --git a/backend/infra/rag/embed/mock_embedder.go b/backend/services/rag/embed/mock_embedder.go
similarity index 100%
rename from backend/infra/rag/embed/mock_embedder.go
rename to backend/services/rag/embed/mock_embedder.go
diff --git a/backend/infra/rag/factory.go b/backend/services/rag/factory.go
similarity index 91%
rename from backend/infra/rag/factory.go
rename to backend/services/rag/factory.go
index c1e8eb0..227c711 100644
--- a/backend/infra/rag/factory.go
+++ b/backend/services/rag/factory.go
@@ -7,12 +7,12 @@ import (
 	"os"
 	"strings"
 
-	ragchunk "github.com/LoveLosita/smartflow/backend/infra/rag/chunk"
-	ragconfig "github.com/LoveLosita/smartflow/backend/infra/rag/config"
-	"github.com/LoveLosita/smartflow/backend/infra/rag/core"
-	ragembed "github.com/LoveLosita/smartflow/backend/infra/rag/embed"
-	ragrerank "github.com/LoveLosita/smartflow/backend/infra/rag/rerank"
-	ragstore "github.com/LoveLosita/smartflow/backend/infra/rag/store"
+	ragchunk "github.com/LoveLosita/smartflow/backend/services/rag/chunk"
+	ragconfig "github.com/LoveLosita/smartflow/backend/services/rag/config"
+	"github.com/LoveLosita/smartflow/backend/services/rag/core"
+	ragembed "github.com/LoveLosita/smartflow/backend/services/rag/embed"
+	ragrerank "github.com/LoveLosita/smartflow/backend/services/rag/rerank"
+	ragstore "github.com/LoveLosita/smartflow/backend/services/rag/store"
 )
 
 // FactoryDeps 描述 Runtime 工厂所需的可选依赖。
diff --git a/backend/infra/rag/observe.go b/backend/services/rag/observe.go
similarity index 92%
rename from backend/infra/rag/observe.go
rename to backend/services/rag/observe.go
index e0d0ddb..97f7ed2 100644
--- a/backend/infra/rag/observe.go
+++ b/backend/services/rag/observe.go
@@ -3,7 +3,7 @@ package rag
 import (
 	"log"
 
-	"github.com/LoveLosita/smartflow/backend/infra/rag/core"
+	"github.com/LoveLosita/smartflow/backend/services/rag/core"
 )
 
 // ObserveLevel 对外暴露统一观测等级别名，避免启动层直接依赖 core 细节。
diff --git a/backend/infra/rag/rag.go b/backend/services/rag/rag.go
similarity index 60%
rename from backend/infra/rag/rag.go
rename to backend/services/rag/rag.go
index c1c3283..25db141 100644
--- a/backend/infra/rag/rag.go
+++ b/backend/services/rag/rag.go
@@ -1,11 +1,11 @@
 package rag
 
 import (
-	"github.com/LoveLosita/smartflow/backend/infra/rag/chunk"
-	"github.com/LoveLosita/smartflow/backend/infra/rag/core"
-	"github.com/LoveLosita/smartflow/backend/infra/rag/embed"
-	"github.com/LoveLosita/smartflow/backend/infra/rag/rerank"
-	"github.com/LoveLosita/smartflow/backend/infra/rag/store"
+	"github.com/LoveLosita/smartflow/backend/services/rag/chunk"
+	"github.com/LoveLosita/smartflow/backend/services/rag/core"
+	"github.com/LoveLosita/smartflow/backend/services/rag/embed"
+	"github.com/LoveLosita/smartflow/backend/services/rag/rerank"
+	"github.com/LoveLosita/smartflow/backend/services/rag/store"
 )
 
 // NewDefaultPipeline 构造默认可运行的 RAG Pipeline。
diff --git a/backend/infra/rag/rerank/eino_reranker.go b/backend/services/rag/rerank/eino_reranker.go
similarity index 86%
rename from backend/infra/rag/rerank/eino_reranker.go
rename to backend/services/rag/rerank/eino_reranker.go
index 931a246..6926eb8 100644
--- a/backend/infra/rag/rerank/eino_reranker.go
+++ b/backend/services/rag/rerank/eino_reranker.go
@@ -4,7 +4,7 @@ import (
 	"context"
 	"errors"
 
-	"github.com/LoveLosita/smartflow/backend/infra/rag/core"
+	"github.com/LoveLosita/smartflow/backend/services/rag/core"
 )
 
 // EinoReranker 是 Eino 重排器占位实现。
diff --git a/backend/infra/rag/rerank/noop_reranker.go b/backend/services/rag/rerank/noop_reranker.go
similarity index 91%
rename from backend/infra/rag/rerank/noop_reranker.go
rename to backend/services/rag/rerank/noop_reranker.go
index 2706873..22485d6 100644
--- a/backend/infra/rag/rerank/noop_reranker.go
+++ b/backend/services/rag/rerank/noop_reranker.go
@@ -4,7 +4,7 @@ import (
 	"context"
 	"sort"
 
-	"github.com/LoveLosita/smartflow/backend/infra/rag/core"
+	"github.com/LoveLosita/smartflow/backend/services/rag/core"
 )
 
 // NoopReranker 是默认重排器（仅按原 score 排序）。
diff --git a/backend/infra/rag/retrieve/vector_retriever.go b/backend/services/rag/retrieve/vector_retriever.go
similarity index 97%
rename from backend/infra/rag/retrieve/vector_retriever.go
rename to backend/services/rag/retrieve/vector_retriever.go
index 0d00950..7ff31e5 100644
--- a/backend/infra/rag/retrieve/vector_retriever.go
+++ b/backend/services/rag/retrieve/vector_retriever.go
@@ -5,7 +5,7 @@ import (
 	"fmt"
 	"strings"
 
-	"github.com/LoveLosita/smartflow/backend/infra/rag/core"
+	"github.com/LoveLosita/smartflow/backend/services/rag/core"
 )
 
 // VectorRetriever 是通用检索器（embed + vector search）。
diff --git a/backend/infra/rag/runtime.go b/backend/services/rag/runtime.go
similarity index 97%
rename from backend/infra/rag/runtime.go
rename to backend/services/rag/runtime.go
index 41eaaad..7c77a21 100644
--- a/backend/infra/rag/runtime.go
+++ b/backend/services/rag/runtime.go
@@ -7,9 +7,9 @@ import (
 	"strings"
 	"time"
 
-	ragconfig "github.com/LoveLosita/smartflow/backend/infra/rag/config"
-	"github.com/LoveLosita/smartflow/backend/infra/rag/core"
-	"github.com/LoveLosita/smartflow/backend/infra/rag/corpus"
+	ragconfig "github.com/LoveLosita/smartflow/backend/services/rag/config"
+	"github.com/LoveLosita/smartflow/backend/services/rag/core"
+	"github.com/LoveLosita/smartflow/backend/services/rag/corpus"
 )
 
 type runtime struct {
@@ -343,7 +343,7 @@ func (r *runtime) recoverPublicPanic(
 		return
 	}
 
-	// 1. runtime 是 RAG Infra 对业务侧暴露的最终方法面，任何下层 panic 都不应再穿透到业务协程。
+	// 1. runtime 是 RAG service 对业务侧暴露的最终方法面，任何下层 panic 都不应再穿透到业务协程。
 	// 2. 这里统一把 panic 转成 error，并补一条结构化观测，方便继续排查是哪一层依赖失控。
 	// 3. 保留 stack 是为了在“进程不崩”的前提下仍能定位根因，避免只剩一句 recovered 无法复盘。
 	panicErr := fmt.Errorf("rag runtime panic recovered: corpus=%s operation=%s panic=%v", corpusName, operation, recovered)
diff --git a/backend/services/rag/service.go b/backend/services/rag/service.go
new file mode 100644
index 0000000..fc38f79
--- /dev/null
+++ b/backend/services/rag/service.go
@@ -0,0 +1,111 @@
+package rag
+
+import (
+	"context"
+
+	ragconfig "github.com/LoveLosita/smartflow/backend/services/rag/config"
+)
+
+// Options 描述 rag-service 需要持有的底层运行时。
+type Options struct {
+	Runtime Runtime
+}
+
+// Service 是 rag-service 对外暴露的统一入口。
+//
+// 职责边界：
+// 1. 负责持有运行时，并把 memory / web 两条能力线统一收口到服务层。
+// 2. 负责在服务入口内完成基于配置的运行时装配。
+// 3. 不直接承载 chunk / embed / store 的实现细节，这些细节下沉到服务树内部子包。
+type Service struct {
+	runtime Runtime
+}
+
+// New 使用调用方传入的运行时构造服务。
+func New(opts Options) *Service {
+	return &Service{runtime: opts.Runtime}
+}
+
+// NewFromConfig 基于服务树内的配置与工厂能力构造自给自足的 RAG 服务。
+func NewFromConfig(ctx context.Context, cfg ragconfig.Config, deps FactoryDeps) (*Service, error) {
+	if !cfg.Enabled {
+		return New(Options{}), nil
+	}
+	runtime, err := NewRuntimeFromConfig(ctx, cfg, deps)
+	if err != nil {
+		return nil, err
+	}
+	return NewWithRuntime(runtime), nil
+}
+
+// Runtime 返回当前服务持有的运行时。
+func (s *Service) Runtime() Runtime {
+	if s == nil {
+		return nil
+	}
+	return s.runtime
+}
+
+// IngestMemory 写入记忆语料。
+func (s *Service) IngestMemory(ctx context.Context, req MemoryIngestRequest) (*IngestResult, error) {
+	if s == nil || s.runtime == nil {
+		return nil, nil
+	}
+	return s.runtime.IngestMemory(ctx, req)
+}
+
+// RetrieveMemory 检索记忆语料。
+func (s *Service) RetrieveMemory(ctx context.Context, req MemoryRetrieveRequest) (*RetrieveResult, error) {
+	if s == nil || s.runtime == nil {
+		return nil, nil
+	}
+	return s.runtime.RetrieveMemory(ctx, req)
+}
+
+// DeleteMemory 删除指定记忆文档。
+func (s *Service) DeleteMemory(ctx context.Context, documentIDs []string) error {
+	if s == nil || s.runtime == nil {
+		return nil
+	}
+	if ctx == nil {
+		ctx = context.Background()
+	}
+	return s.runtime.DeleteMemory(ctx, documentIDs)
+}
+
+// IngestWeb 写入网页语料。
+func (s *Service) IngestWeb(ctx context.Context, req WebIngestRequest) (*IngestResult, error) {
+	if s == nil || s.runtime == nil {
+		return nil, nil
+	}
+	return s.runtime.IngestWeb(ctx, req)
+}
+
+// RetrieveWeb 检索网页语料。
+func (s *Service) RetrieveWeb(ctx context.Context, req WebRetrieveRequest) (*RetrieveResult, error) {
+	if s == nil || s.runtime == nil {
+		return nil, nil
+	}
+	return s.runtime.RetrieveWeb(ctx, req)
+}
+
+// EnsureRuntime 返回一个可继续向下传递的运行时引用。
+func (s *Service) EnsureRuntime() Runtime {
+	if s == nil {
+		return nil
+	}
+	return s.runtime
+}
+
+// SetRuntime 允许在装配阶段延迟注入运行时。
+func (s *Service) SetRuntime(runtime Runtime) {
+	if s == nil {
+		return
+	}
+	s.runtime = runtime
+}
+
+// NewWithRuntime 用显式运行时构造服务。
+func NewWithRuntime(runtime Runtime) *Service {
+	return New(Options{Runtime: runtime})
+}
diff --git a/backend/infra/rag/store/inmemory_store.go b/backend/services/rag/store/inmemory_store.go
similarity index 98%
rename from backend/infra/rag/store/inmemory_store.go
rename to backend/services/rag/store/inmemory_store.go
index f15ba28..19e4df9 100644
--- a/backend/infra/rag/store/inmemory_store.go
+++ b/backend/services/rag/store/inmemory_store.go
@@ -10,7 +10,7 @@ import (
 	"sync"
 	"time"
 
-	"github.com/LoveLosita/smartflow/backend/infra/rag/core"
+	"github.com/LoveLosita/smartflow/backend/services/rag/core"
 )
 
 // InMemoryVectorStore 是本地开发用向量存储实现。
diff --git a/backend/infra/rag/store/milvus_store.go b/backend/services/rag/store/milvus_store.go
similarity index 99%
rename from backend/infra/rag/store/milvus_store.go
rename to backend/services/rag/store/milvus_store.go
index 2e8db58..39a92a6 100644
--- a/backend/infra/rag/store/milvus_store.go
+++ b/backend/services/rag/store/milvus_store.go
@@ -14,7 +14,7 @@ import (
 	"sync"
 	"time"
 
-	"github.com/LoveLosita/smartflow/backend/infra/rag/core"
+	"github.com/LoveLosita/smartflow/backend/services/rag/core"
 )
 
 // MilvusConfig 描述 Milvus REST 存储配置。
diff --git a/backend/infra/rag/store/vector_store.go b/backend/services/rag/store/vector_store.go
similarity index 75%
rename from backend/infra/rag/store/vector_store.go
rename to backend/services/rag/store/vector_store.go
index 3535fdb..5549461 100644
--- a/backend/infra/rag/store/vector_store.go
+++ b/backend/services/rag/store/vector_store.go
@@ -1,6 +1,6 @@
 package store
 
-import "github.com/LoveLosita/smartflow/backend/infra/rag/core"
+import "github.com/LoveLosita/smartflow/backend/services/rag/core"
 
 // EnsureCompile 用于静态校验实现是否满足接口。
 func EnsureCompile() {
diff --git a/docs/backend/微服务四步迁移与第二阶段并行开发计划.md b/docs/backend/微服务四步迁移与第二阶段并行开发计划.md
index 97e4bed..62f6ac1 100644
--- a/docs/backend/微服务四步迁移与第二阶段并行开发计划.md
+++ b/docs/backend/微服务四步迁移与第二阶段并行开发计划.md
@@ -36,6 +36,8 @@
 4. 消费侧已经按服务 consumer group 隔离，不再用一个 worker 吃全部事件。
 5. 当前仍是单体进程内多 worker 装配；worker 后续会跟随对应服务一起迁出，不在阶段 1 直接拆进程。
 
+阶段 1.5 / 1.6 也已经先落地完毕：`backend/services/llm` 和 `backend/services/rag` 已经成为当前 canonical 入口，`backend/infra/llm` 和 `backend/infra/rag` 的 `.go` 旧实现已删除，仅保留迁移说明文档。当前仍然是单体进程内多 worker 装配，llm / rag 先完成服务化收口，还没有进入 gozero 进程拆分。
+
 所以后续路线不是再补一次 outbox 基建，而是在这个阶段 1 基线上，按服务边界逐个把 gozero 服务、DAO / model / worker 和启动入口迁出去。
 
 ---
@@ -78,6 +80,8 @@ gozero 服务负责领域能力：
 > 说明：`agent` 和 `memory` 都可以单独成服务，不应再被写成“公共能力”；其中 `agent` 更像对外对话编排服务，`memory` 更像其支撑服务/worker 服务。
 >
 > 说明：`llm-service` 先抽成全仓统一模型出口，`rag-service` 再抽成检索基础设施服务；`rag-service` 只能依赖 `llm-service`，不反向依赖具体业务服务。
+>
+> 当前状态：`llm-service` / `rag-service` 这两个边界已经先做成 `backend/services/*` 的服务内模块，调用仍由 `backend/cmd/start.go` 在同一进程内装配，不是 gozero 独立进程。
 
 ### 3.3 事件层
 
@@ -118,8 +122,8 @@ gozero 服务负责领域能力：
 | --- | --- | --- | --- |
 | 0 | 语义冻结和基线确认（已完成） | 阶段 0 已作为历史基线保存；后续只在契约变化时回看 | `go test ./...`，`api / worker / all` 启动 smoke |
 | 1 | Outbox v2 基建（已完成，当前基线） | 当前已具备阶段 1 保存点：服务级 outbox 表、topic、group 和多 worker 装配已打通 | 已完成健康检查、服务级 outbox 写入/投递/消费 smoke、Kafka group lag 核对 |
-| 1.5 | 先抽 llm-service | 统一模型调用、provider 路由、流式输出和审计后 commit | course / active-scheduler / memory 模型调用 smoke |
-| 1.6 | 再抽 rag-service | 向量化、召回、重排、检索能力跑通后 commit | memory retrieve / rerank smoke |
+| 1.5 | 先抽 llm-service（已完成） | 已完成，`backend/services/llm` 作为当前 canonical 入口 | `go test ./...` + course / active-scheduler / memory 模型调用 smoke |
+| 1.6 | 再抽 rag-service（已完成） | 已完成，`backend/services/rag` 作为当前 canonical 入口 | `go test ./...` + memory retrieve / rerank smoke |
 | 2 | 先拆 user/auth | user 路由、JWT 签发和 token 额度治理独立后 commit | 注册/登录/刷新/登出 smoke + token quota 回归 |
 | 3 | 再拆 notification | notification 服务能独立消费和重试后 commit | notification E2E smoke + worker-only smoke |
 | 4 | 再拆 active-scheduler | 预览生成和确认链路通过 gozero 服务跑通后 commit | dry-run / preview / confirm smoke |
@@ -229,7 +233,7 @@ flowchart LR
 
 ---
 
-### 4.4 阶段 1.5：先抽 llm-service
+### 4.4 阶段 1.5：先抽 llm-service（已完成）
 
 目标：
 
@@ -237,6 +241,12 @@ flowchart LR
 2. 让 `course`、`active-scheduler`、`memory`、`agent` 对模型调用的依赖先收口到统一服务。
 3. 先把模型 provider 路由、流式输出、限流、审计这些共性收束起来，避免每个服务各写一份。
 
+当前状态：
+
+1. 代码已经落到 `backend/services/llm`。
+2. `backend/infra/llm` 的 `.go` 旧实现已删除，仅保留迁移说明。
+3. 仍由 `backend/cmd/start.go` 在同一进程内装配，尚未引入 gozero 独立服务进程。
+
 这一步要做的事：
 
 1. 把当前分散在业务服务里的模型调用入口改成统一调用 `llm-service`。
@@ -260,7 +270,7 @@ flowchart LR
 
 ---
 
-### 4.5 阶段 1.6：再抽 rag-service
+### 4.5 阶段 1.6：再抽 rag-service（已完成）
 
 目标：
 
@@ -268,6 +278,12 @@ flowchart LR
 2. 让向量化、召回、重排、向量库读写先进入独立服务。
 3. 明确 `rag-service` 只能依赖 `llm-service` 做 embedding / rerank，不反向依赖业务服务。
 
+当前状态：
+
+1. 代码已经落到 `backend/services/rag`。
+2. `backend/infra/rag` 的 `.go` 旧实现已删除，仅保留迁移说明。
+3. 仍由 `backend/cmd/start.go` 在同一进程内装配，尚未引入 gozero 独立服务进程。
+
 这一步要做的事：
 
 1. 把当前分散在 `memory`、`agent` 里的检索逻辑改成统一调用 `rag-service`。
@@ -474,6 +490,8 @@ flowchart LR
 
 当前建议按这个顺序推进：
 
+注：阶段 1.5 / 1.6 已完成，当前实际推进可从阶段 2 开始。
+
 1. 以阶段 1 的服务级 outbox 为当前基线，不再回头做共享 outbox 方案。
 2. 先切 llm-service，把统一模型出口从各业务服务里抽出去。
 3. 再切 rag-service，把检索基础设施从 memory / agent 里抽出去。