feat: filter meaningful conversations with local model

CHANGELOG.md

@@ -4,6 +4,10 @@
 
 ## 未发布
 
+- 聊天总结改为大模型两级内容价值判断：先判断是否值得保留，再判断是否值得形成故事；纯寒暄不再生成摘要或故事。
+- 新增内容价值判断记录，保留模型、提示词版本、置信度和舍弃原因，支持任务幂等与隐私删除。
+- 本地开发新增 GPU Ollama/Qwen2.5 7B 部署方式，并通过大模型管理的测试、激活和 worker 生成链路完成端到端验证。
+- 修复大模型“测试连接”更新语句和审计结果枚举错误，测试通过的模型现在可以正常激活。
 - 维护端默认人设升级为老人陪伴对话规范，覆盖倾听、单问题追问、沉默与重复处理、事实边界、敏感话题和医疗安全。
 - 家庭端长者档案新增日常陪伴、回忆往事、兴趣聊天三个模板，并补充兴趣、优先话题和避免话题的创建入口。
 - Agent固定评测加强单问题追问、先回应再追问、沉默不施压、尊重重复表达和情绪优先回应检查。

compose.yaml

@@ -155,6 +155,26 @@ services:
     volumes:
       - ./deploy/keycloak:/opt/keycloak/data/import:ro
 
+  ollama:
+    profiles: [local-model]
+    image: ollama/ollama:latest
+    ports:
+      - "${OLLAMA_PORT:-11434}:11434"
+    volumes:
+      - ollama-data:/root/.ollama
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+    healthcheck:
+      test: ["CMD", "ollama", "list"]
+      interval: 10s
+      timeout: 5s
+      retries: 30
+
   otel-collector:
     image: otel/opentelemetry-collector-contrib:latest
     command: ["--config=/etc/otelcol-contrib/config.yaml"]
@@ -325,3 +345,4 @@ volumes:
   postgres-data:
   redis-data:
   minio-data:
+  ollama-data:

deploy/local-model/README.md

@@ -0,0 +1,31 @@
+# 本地聊天总结模型
+
+本地开发环境使用 Ollama 提供 OpenAI 兼容接口，默认验证模型为 `qwen2.5:7b`。模型只监听本机 `11434` 端口，应用容器通过 `http://ollama:11434/v1` 访问。
+
+## 启动与下载模型
+
+```powershell
+docker compose --profile local-model up -d ollama
+docker compose --profile local-model exec ollama ollama pull qwen2.5:7b
+```
+
+随后在客户门户的“大模型管理”中创建并激活配置：
+
+- 服务类型：`OpenAI 兼容接口`
+- 接口基础地址：`http://ollama:11434/v1`
+- 模型 ID：`qwen2.5:7b`
+- API Key：本地 Ollama 不校验密钥，可填写长度至少 8 位的本地占位值
+- 建议温度：`0.1`
+- 建议最大输出：`1200`
+
+必须先执行“测试连接”，通过后才能“设为当前模型”。配置激活后，worker 会在每个总结任务开始时读取当前模型，不需要重启。
+
+## 真实模型回归测试
+
+```powershell
+$env:LOCAL_MODEL_BASE_URL='http://localhost:11434/v1'
+$env:LOCAL_MODEL_ID='qwen2.5:7b'
+go test ./internal/extraction -run TestLocalModelContentValueDecisions -count=1 -v
+```
+
+测试覆盖三类输入：纯寒暄应舍弃、有用近况只生成摘要、完整个人回忆生成摘要和故事候选。

docs/device-api-compatibility.md

@@ -1,7 +1,7 @@
 # 设备 API 兼容性约定
 
 > 当前设备接口主版本：`/device/v2`  
-> 当前接口契约：OpenAPI 0.70.0
+> 当前接口契约：OpenAPI 0.73.0
 > 适用对象：云端 API、终端固件、模拟终端和设备联调工具
 
 ## 1. 兼容目标

internal/contentmodel/service.go

@@ -209,15 +209,19 @@ func (s *Service) Test(ctx context.Context, principal identity.Principal, tenant
 		testStatus, errorCode = "failed", normalizedErrorCode(err)
 	}
 	var result Model
+	auditResult := "success"
+	if testStatus != "passed" {
+		auditResult = "failure"
+	}
 	updateErr := database.Transact(ctx, s.registry.pool, func(tx pgx.Tx) error {
-		row := tx.QueryRow(ctx, modelUpdateReturning+` WHERE cm.tenant_id=$1::uuid AND cm.public_id=$2::uuid`, current.tenantID, publicID, testStatus, errorCode, current.userID, current.tenantPublicID)
+		row := tx.QueryRow(ctx, modelUpdateReturning, current.tenantID, publicID, testStatus, errorCode, current.userID, current.tenantPublicID)
 		updated, scanErr := scanModel(row)
 		if scanErr != nil {
 			return scanErr
 		}
 		result = updated.Model
 		resourceID := result.ID
-		return s.audit.Record(ctx, tx, audit.Event{TenantID: &current.tenantID, ActorType: "user", ActorPublicID: &current.userPublicID, Action: "content_model.test", ResourceType: "content_model", ResourceID: &resourceID, RequestID: requestID, Purpose: "验证总结模型连通性", Result: testStatus, Detail: map[string]any{"error_code": errorCode}})
+		return s.audit.Record(ctx, tx, audit.Event{TenantID: &current.tenantID, ActorType: "user", ActorPublicID: &current.userPublicID, Action: "content_model.test", ResourceType: "content_model", ResourceID: &resourceID, RequestID: requestID, Purpose: "验证总结模型连通性", Result: auditResult, Detail: map[string]any{"test_status": testStatus, "error_code": errorCode}})
 	})
 	if updateErr != nil {
 		return Model{}, updateErr
@@ -359,13 +363,18 @@ func normalizeInput(input *CreateInput) {
 func validateInput(input CreateInput) error {
 	parsed, err := url.Parse(input.BaseURL)
 	validProvider := input.Provider == "ark" || input.Provider == "openai_compatible"
-	validURL := err == nil && parsed.Host != "" && (parsed.Scheme == "https" || (parsed.Scheme == "http" && (parsed.Hostname() == "localhost" || parsed.Hostname() == "127.0.0.1")))
+	validURL := err == nil && parsed.Host != "" && (parsed.Scheme == "https" || (parsed.Scheme == "http" && isApprovedLocalModelHost(parsed.Hostname())))
 	if !validUUID(input.TenantID) || len([]rune(input.DisplayName)) < 1 || len([]rune(input.DisplayName)) > 120 || !validProvider || !validURL || input.ModelID == "" || len(input.APIKey) < 8 || input.Temperature < 0 || input.Temperature > 2 || input.MaxTokens < 256 || input.MaxTokens > 8192 {
 		return ErrInvalid
 	}
 	return nil
 }
 
+func isApprovedLocalModelHost(host string) bool {
+	host = strings.ToLower(strings.TrimSpace(host))
+	return host == "localhost" || host == "127.0.0.1" || host == "::1" || host == "ollama"
+}
+
 func credentialHint(key string) string {
 	runes := []rune(strings.TrimSpace(key))
 	if len(runes) <= 4 {
@@ -398,7 +407,7 @@ func validUUID(value string) bool { _, err := uuid.Parse(value); return err == n
 
 const modelSelect = `SELECT cm.public_id::text,t.public_id::text,cm.display_name,cm.provider,cm.base_url,cm.model_id,cm.credential_hint,cm.temperature::float8,cm.max_tokens,cm.status,cm.test_status,cm.last_tested_at,COALESCE(cm.last_error_code,''),cm.activated_at,cm.version,cm.created_at,cm.updated_at,cm.tenant_id::text,cm.encrypted_api_key,cm.nonce,cm.key_version FROM content_model_config cm JOIN tenant t ON t.id=cm.tenant_id`
 const modelReturning = `cm.public_id::text,$4::text,cm.display_name,cm.provider,cm.base_url,cm.model_id,cm.credential_hint,cm.temperature::float8,cm.max_tokens,cm.status,cm.test_status,cm.last_tested_at,COALESCE(cm.last_error_code,''),cm.activated_at,cm.version,cm.created_at,cm.updated_at,cm.tenant_id::text,cm.encrypted_api_key,cm.nonce,cm.key_version`
-const modelUpdateReturning = `UPDATE content_model_config cm SET test_status=$3,last_tested_at=now(),last_error_code=NULLIF($4,''),updated_at=now(),version=version+1,updated_by=$5::uuid RETURNING cm.public_id::text,$6::text,cm.display_name,cm.provider,cm.base_url,cm.model_id,cm.credential_hint,cm.temperature::float8,cm.max_tokens,cm.status,cm.test_status,cm.last_tested_at,COALESCE(cm.last_error_code,''),cm.activated_at,cm.version,cm.created_at,cm.updated_at,cm.tenant_id::text,cm.encrypted_api_key,cm.nonce,cm.key_version`
+const modelUpdateReturning = `UPDATE content_model_config cm SET test_status=$3,last_tested_at=now(),last_error_code=NULLIF($4,''),updated_at=now(),version=version+1,updated_by=$5::uuid WHERE cm.tenant_id=$1::uuid AND cm.public_id=$2::uuid RETURNING cm.public_id::text,$6::text,cm.display_name,cm.provider,cm.base_url,cm.model_id,cm.credential_hint,cm.temperature::float8,cm.max_tokens,cm.status,cm.test_status,cm.last_tested_at,COALESCE(cm.last_error_code,''),cm.activated_at,cm.version,cm.created_at,cm.updated_at,cm.tenant_id::text,cm.encrypted_api_key,cm.nonce,cm.key_version`
 
 type rowScanner interface{ Scan(...any) error }
 

internal/contentmodel/service_integration_test.go

@@ -0,0 +1,67 @@
+package contentmodel
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"testing"
+
+	"data-terminal-cloud/internal/audit"
+	"data-terminal-cloud/internal/extraction"
+	"data-terminal-cloud/internal/identity"
+
+	"github.com/jackc/pgx/v5/pgxpool"
+)
+
+func TestServiceTestsAndActivatesCompatibleModel(t *testing.T) {
+	databaseURL := os.Getenv("TEST_DATABASE_URL")
+	if databaseURL == "" {
+		t.Skip("TEST_DATABASE_URL is not set")
+	}
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		_ = json.NewEncoder(w).Encode(map[string]any{"choices": []any{map[string]any{"message": map[string]any{"content": `{"meaningful":false,"story_worthy":false,"decision_reason":"连通性测试内容不形成故事","summary":"","keywords":[],"open_topic":"","story_title":"","story_content":"","fact_content":"","confidence":0.99}`}}}})
+	}))
+	defer server.Close()
+	ctx := context.Background()
+	pool, err := pgxpool.New(ctx, databaseURL)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer pool.Close()
+	registry, err := NewRegistry(pool, "1f1e1d1c1b1a191817161514131211100f0e0d0c0b0a09080706050403020100", extraction.DeterministicGenerator{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	service, err := NewService(registry, audit.NewRepository())
+	if err != nil {
+		t.Fatal(err)
+	}
+	principal := identity.Principal{Issuer: "http://localhost:58180/realms/voice-local", Subject: "00000000-0000-0000-0000-000000000101"}
+	created, err := service.Create(ctx, principal, CreateInput{
+		TenantID: "10000000-0000-0000-0000-000000000010", DisplayName: "集成测试模型", Provider: "openai_compatible",
+		BaseURL: server.URL + "/v1", ModelID: "test-model", APIKey: "integration-key", Temperature: 0.1, MaxTokens: 512,
+	}, "content-model-integration-create")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer func() {
+		_, _ = pool.Exec(ctx, `DELETE FROM audit_event WHERE resource_id=$1::uuid`, created.ID)
+		_, _ = pool.Exec(ctx, `DELETE FROM content_model_config WHERE public_id=$1::uuid`, created.ID)
+	}()
+	tested, err := service.Test(ctx, principal, created.TenantID, created.ID, "content-model-integration-test")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if tested.TestStatus != "passed" || tested.Version != created.Version+1 {
+		t.Fatalf("tested model=%+v", tested)
+	}
+	activated, err := service.Activate(ctx, principal, created.TenantID, created.ID, "content-model-integration-activate", tested.Version)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if activated.Status != "active" || activated.Version != tested.Version+1 {
+		t.Fatalf("activated model=%+v", activated)
+	}
+}

internal/contentmodel/service_test.go

@@ -0,0 +1,25 @@
+package contentmodel
+
+import "testing"
+
+func TestValidateInputAllowsComposeOllamaEndpoint(t *testing.T) {
+	input := CreateInput{
+		TenantID: "10000000-0000-0000-0000-000000000001", DisplayName: "本地 Qwen",
+		Provider: "openai_compatible", BaseURL: "http://ollama:11434/v1", ModelID: "qwen2.5:7b",
+		APIKey: "ollama-local-key", Temperature: 0.1, MaxTokens: 1200,
+	}
+	if err := validateInput(input); err != nil {
+		t.Fatalf("compose Ollama endpoint was rejected: %v", err)
+	}
+}
+
+func TestValidateInputRejectsUnapprovedPlainHTTPEndpoint(t *testing.T) {
+	input := CreateInput{
+		TenantID: "10000000-0000-0000-0000-000000000001", DisplayName: "不安全端点",
+		Provider: "openai_compatible", BaseURL: "http://metadata.internal/v1", ModelID: "model",
+		APIKey: "test-api-key", Temperature: 0.1, MaxTokens: 1200,
+	}
+	if err := validateInput(input); err == nil {
+		t.Fatal("unapproved plaintext endpoint was accepted")
+	}
+}

internal/extraction/generator.go

@@ -23,16 +23,19 @@ type GenerationInput struct {
 }
 
 type GenerationResult struct {
-	Summary       string   `json:"summary"`
+	Meaningful     bool     `json:"meaningful"`
-	OpenTopic     string   `json:"open_topic"`
+	StoryWorthy    bool     `json:"story_worthy"`
-	StoryTitle    string   `json:"story_title"`
+	DecisionReason string   `json:"decision_reason"`
-	StoryContent  string   `json:"story_content"`
+	Summary        string   `json:"summary"`
-	FactContent   string   `json:"fact_content"`
+	OpenTopic      string   `json:"open_topic"`
-	Keywords      []string `json:"keywords"`
+	StoryTitle     string   `json:"story_title"`
-	Confidence    float64  `json:"confidence"`
+	StoryContent   string   `json:"story_content"`
-	ModelRef      string   `json:"-"`
+	FactContent    string   `json:"fact_content"`
-	PromptVersion string   `json:"-"`
+	Keywords       []string `json:"keywords"`
-	AllowStory    bool     `json:"-"`
+	Confidence     float64  `json:"confidence"`
+	ModelRef       string   `json:"-"`
+	PromptVersion  string   `json:"-"`
+	AllowStory     bool     `json:"-"`
 }
 
 type Generator interface {
@@ -46,17 +49,33 @@ func (DeterministicGenerator) Generate(_ context.Context, input GenerationInput)
 	if content == "" {
 		return GenerationResult{}, errors.New("extraction content is required")
 	}
+	meaningful, storyWorthy, reason := deterministicContentDecision(content)
+	if !meaningful {
+		return GenerationResult{
+			Meaningful: false, StoryWorthy: false, DecisionReason: reason,
+			Confidence: 0.9, ModelRef: "local-deterministic-extractor", PromptVersion: "content-value-v1",
+		}, nil
+	}
+	storyTitle, storyContent, factContent := "", "", ""
+	if storyWorthy {
+		storyTitle = "关于“" + truncateRunes(content, 24) + "”的回忆"
+		storyContent = content
+		factContent = content
+	}
 	return GenerationResult{
-		Summary:       "长者在本次通话中谈到：" + content,
+		Meaningful:     true,
-		OpenTopic:     "下次可以继续询问长者刚才提到的经历，并让长者自行补充人物、时间和地点。",
+		StoryWorthy:    storyWorthy,
-		StoryTitle:    "关于“" + truncateRunes(content, 24) + "”的回忆",
+		DecisionReason: reason,
-		StoryContent:  content,
+		Summary:        "长者在本次通话中谈到：" + content,
-		FactContent:   content,
+		OpenTopic:      "下次可以继续询问长者刚才提到的经历，并让长者自行补充人物、时间和地点。",
-		Keywords:      deterministicKeywords(content),
+		StoryTitle:     storyTitle,
-		Confidence:    0.9,
+		StoryContent:   storyContent,
-		ModelRef:      "local-deterministic-extractor",
+		FactContent:    factContent,
-		PromptVersion: "story-v1",
+		Keywords:       deterministicKeywords(content),
-		AllowStory:    true,
+		Confidence:     0.9,
+		ModelRef:       "local-deterministic-extractor",
+		PromptVersion:  "content-value-deterministic-v2",
+		AllowStory:     storyWorthy,
 	}, nil
 }
 
@@ -67,12 +86,15 @@ func sensitiveGeneration(decision contentpolicy.Decision) GenerationResult {
 	}
 	reason := safePolicyToken(decision.Reason)
 	return GenerationResult{
-		Summary:       "本次通话包含需要人工谨慎处理的敏感信息，系统未生成普通故事卡片。",
+		Meaningful:     true,
-		OpenTopic:     "如确有必要处理本次敏感内容，应由有权限人员在受控流程中单独核对。",
+		StoryWorthy:    false,
-		Confidence:    1,
+		DecisionReason: "safety_policy_blocked_story",
-		ModelRef:      modelRef,
+		Summary:        "本次通话包含需要人工谨慎处理的敏感信息，系统未生成普通故事卡片。",
-		PromptVersion: "summary-sensitive-filter-v1:" + reason,
+		OpenTopic:      "如确有必要处理本次敏感内容，应由有权限人员在受控流程中单独核对。",
-		AllowStory:    false,
+		Confidence:     1,
+		ModelRef:       modelRef,
+		PromptVersion:  "summary-sensitive-filter-v1:" + reason,
+		AllowStory:     false,
 	}
 }
 
@@ -133,7 +155,7 @@ func (g *ArkGenerator) Generate(ctx context.Context, input GenerationInput) (Gen
 	requestBody := map[string]any{
 		"model": g.model,
 		"messages": []map[string]string{
-			{"role": "system", "content": "你是长者口述内容整理助手。只依据输入字幕生成可人工核对的候选，不补充未出现的事实，不把推断写成事实。必须返回JSON对象，字段为summary、keywords、open_topic、story_title、story_content、fact_content、confidence。summary概括本次内容，keywords是1至8个简短关键词字符串组成的数组，open_topic给出一个温和的后续话题，story和fact保留原意，confidence取0到1。"},
+			{"role": "system", "content": contentValuePrompt},
 			{"role": "user", "content": content},
 		},
 		"temperature":     g.temperature,
@@ -186,30 +208,75 @@ func (g *ArkGenerator) Generate(ctx context.Context, input GenerationInput) (Gen
 		return GenerationResult{}, outbox.Retryable(fmt.Errorf("decode Ark structured content: %w", err))
 	}
 	result.Summary = strings.TrimSpace(result.Summary)
+	result.DecisionReason = strings.TrimSpace(result.DecisionReason)
 	result.OpenTopic = strings.TrimSpace(result.OpenTopic)
 	result.StoryTitle = strings.TrimSpace(result.StoryTitle)
 	result.StoryContent = strings.TrimSpace(result.StoryContent)
 	result.FactContent = strings.TrimSpace(result.FactContent)
 	result.Keywords = normalizedKeywords(result.Keywords)
+	normalizeDecisionContent(&result)
 	if !validGeneratedContent(result) {
 		return GenerationResult{}, outbox.Retryable(errors.New("Ark structured content failed validation"))
 	}
 	result.ModelRef = g.provider + ":" + g.model
-	result.PromptVersion = "story-" + g.provider + "-json-v2"
+	result.PromptVersion = "content-value-" + g.provider + "-json-v3"
-	result.AllowStory = true
+	result.AllowStory = result.Meaningful && result.StoryWorthy
 	return result, nil
 }
 
+func normalizeDecisionContent(result *GenerationResult) {
+	if !result.Meaningful {
+		result.StoryWorthy = false
+		result.Summary = ""
+		result.OpenTopic = ""
+		result.StoryTitle = ""
+		result.StoryContent = ""
+		result.FactContent = ""
+		result.Keywords = nil
+		return
+	}
+	if !result.StoryWorthy {
+		result.StoryTitle = ""
+		result.StoryContent = ""
+		result.FactContent = ""
+	}
+}
+
 func validGeneratedContent(result GenerationResult) bool {
-	return result.Summary != "" && len([]rune(result.Summary)) <= 2000 &&
+	if result.Confidence < 0 || result.Confidence > 1 || result.DecisionReason == "" || len([]rune(result.DecisionReason)) > 500 {
-		result.OpenTopic != "" && len([]rune(result.OpenTopic)) <= 1000 &&
+		return false
-		result.StoryTitle != "" && len([]rune(result.StoryTitle)) <= 300 &&
+	}
+	if !result.Meaningful {
+		return !result.StoryWorthy && result.Summary == "" && result.OpenTopic == "" && result.StoryTitle == "" && result.StoryContent == "" && result.FactContent == "" && len(result.Keywords) == 0
+	}
+	if result.Summary == "" || len([]rune(result.Summary)) > 2000 ||
+		result.OpenTopic == "" || len([]rune(result.OpenTopic)) > 1000 ||
+		len(result.Keywords) < 1 || len(result.Keywords) > 8 {
+		return false
+	}
+	if !result.StoryWorthy {
+		return result.StoryTitle == "" && result.StoryContent == "" && result.FactContent == ""
+	}
+	return result.StoryTitle != "" && len([]rune(result.StoryTitle)) <= 300 &&
 		result.StoryContent != "" && len([]rune(result.StoryContent)) <= 20000 &&
 		result.FactContent != "" && len([]rune(result.FactContent)) <= 5000 &&
-		len(result.Keywords) >= 1 && len(result.Keywords) <= 8 &&
 		result.Confidence >= 0 && result.Confidence <= 1
 }
 
+const contentValuePrompt = `你是长者聊天内容筛选与整理助手。只依据输入对话判断和整理，不补充未出现的事实，不把推断写成事实。
+
+先做两级判断：
+1. meaningful：对话是否包含值得保留的信息。只有问候、寒暄、嗯啊、测试设备、重复句、零散且无法理解的短句，必须判为false并舍弃，不生成摘要。
+2. story_worthy：仅在meaningful为true时判断是否值得形成故事。具体的个人经历或回忆，且人物、时间、地点、事件、行动、感受中至少出现两类，或形成了清楚的事件脉络，才判为true。普通偏好、单个事实或简短近况可以总结，但story_worthy应为false。
+
+必须只返回JSON对象，字段为meaningful、story_worthy、decision_reason、summary、keywords、open_topic、story_title、story_content、fact_content、confidence。
+- decision_reason用一句简短中文说明判断依据。
+- meaningful=false时，除decision_reason和confidence外，其余文本必须为空字符串，keywords必须为空数组，story_worthy必须为false。
+- meaningful=true时，summary概括有意义内容，keywords为1至8个简短关键词，open_topic给出一个温和且不重复追问的后续话题。
+- story_worthy=false时，story_title、story_content、fact_content必须为空字符串。
+- story_worthy=true时，story_title和story_content忠实整理原话，fact_content只写对话中可核对的事实。
+- confidence取0到1。`
+
 func normalizedKeywords(values []string) []string {
 	result := make([]string, 0, min(len(values), 8))
 	seen := make(map[string]struct{}, len(values))
@@ -255,3 +322,32 @@ func deterministicKeywords(content string) []string {
 	}
 	return []string{string(runes)}
 }
+
+func deterministicContentDecision(content string) (meaningful, storyWorthy bool, reason string) {
+	compact := strings.Map(func(character rune) rune {
+		if unicode.IsSpace(character) || strings.ContainsRune("，。！？；、,.!?;：:\"“”‘’（）()-", character) {
+			return -1
+		}
+		return character
+	}, content)
+	for _, label := range []string{"长者", "豆包", "助手", "用户", "agent", "subject"} {
+		compact = strings.ReplaceAll(compact, label, "")
+	}
+	for _, phrase := range []string{"你好", "您好", "在吗", "嗯", "啊", "哦", "好的", "知道了", "没事", "再见", "拜拜", "听得到吗", "能听见吗", "测试"} {
+		compact = strings.ReplaceAll(compact, phrase, "")
+	}
+	if len([]rune(compact)) < 12 {
+		return false, false, "内容仅为寒暄、确认或零散短句"
+	}
+	storySignals := 0
+	for _, phrase := range []string{"小时候", "年轻时", "那年", "当时", "后来", "以前", "记得", "有一次", "我们", "我和", "去了", "发生", "因为", "结果", "高兴", "难过", "想起"} {
+		if strings.Contains(content, phrase) {
+			storySignals++
+		}
+	}
+	storyWorthy = len([]rune(compact)) >= 48 && storySignals >= 2
+	if storyWorthy {
+		return true, true, "包含较完整的个人经历与事件脉络"
+	}
+	return true, false, "包含可保留的信息，但尚不足以形成完整故事"
+}

internal/extraction/generator_local_model_test.go

@@ -0,0 +1,58 @@
+package extraction
+
+import (
+	"context"
+	"os"
+	"testing"
+	"time"
+)
+
+func TestLocalModelContentValueDecisions(t *testing.T) {
+	baseURL := os.Getenv("LOCAL_MODEL_BASE_URL")
+	if baseURL == "" {
+		t.Skip("LOCAL_MODEL_BASE_URL is not set")
+	}
+	modelID := os.Getenv("LOCAL_MODEL_ID")
+	if modelID == "" {
+		modelID = "qwen2.5:7b"
+	}
+	generator, err := NewCompatibleGenerator("openai_compatible", baseURL, "ollama-local-key", modelID, 0.1, 1200, 2*time.Minute)
+	if err != nil {
+		t.Fatal(err)
+	}
+	tests := []struct {
+		name, content          string
+		meaningful, storyWorth bool
+	}{
+		{
+			name:    "discard small talk",
+			content: "豆包：您好，今天过得怎么样？\n长者：还行。\n豆包：吃饭了吗？\n长者：嗯，吃了。\n豆包：那您好好休息。\n长者：好，再见。",
+		},
+		{
+			name:       "keep useful update without story",
+			content:    "豆包：今天胃口怎么样？\n长者：今天胃口挺好，中午吃了一碗西红柿鸡蛋面，晚上想吃点清淡的。",
+			meaningful: true,
+		},
+		{
+			name:       "create story candidate from coherent memory",
+			content:    "豆包：您还记得第一次当老师的情形吗？\n长者：记得。那是1978年秋天，我二十二岁，被分到河边小学。第一天是姐姐骑自行车送我去的，教室里有三十多个孩子。我紧张得把粉笔掉在地上，一个小姑娘帮我捡起来，全班都笑了，我也跟着笑。后来那个小姑娘一直读到大学，很多年后还回来看过我。",
+			meaningful: true, storyWorth: true,
+		},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			ctx, cancel := context.WithTimeout(context.Background(), 2*time.Minute)
+			defer cancel()
+			result, generateErr := generator.Generate(ctx, GenerationInput{Content: test.content})
+			if generateErr != nil {
+				t.Fatal(generateErr)
+			}
+			if result.Meaningful != test.meaningful || result.StoryWorthy != test.storyWorth || result.AllowStory != test.storyWorth {
+				t.Fatalf("decision mismatch: meaningful=%v story_worthy=%v allow_story=%v reason=%q summary=%q", result.Meaningful, result.StoryWorthy, result.AllowStory, result.DecisionReason, result.Summary)
+			}
+			if !result.Meaningful && (result.Summary != "" || result.StoryContent != "") {
+				t.Fatalf("discarded conversation produced content: %+v", result)
+			}
+		})
+	}
+}

internal/extraction/generator_test.go

@@ -20,7 +20,7 @@ func TestArkGeneratorProducesValidatedContent(t *testing.T) {
 		if err := json.NewDecoder(r.Body).Decode(&body); err != nil || body["model"] != "doubao-test" {
 			t.Fatalf("request body=%v err=%v", body, err)
 		}
-		_ = json.NewEncoder(w).Encode(map[string]any{"choices": []any{map[string]any{"message": map[string]any{"content": `{"summary":"谈到河边小学","keywords":["河边小学","童年"],"open_topic":"下次可以聊聊同学","story_title":"河边小学的回忆","story_content":"小时候在河边小学读书。","fact_content":"曾在河边小学读书","confidence":0.82}`}}}})
+		_ = json.NewEncoder(w).Encode(map[string]any{"choices": []any{map[string]any{"message": map[string]any{"content": `{"meaningful":true,"story_worthy":true,"decision_reason":"包含具体的童年求学回忆","summary":"谈到河边小学","keywords":["河边小学","童年"],"open_topic":"下次可以聊聊同学","story_title":"河边小学的回忆","story_content":"小时候在河边小学读书。","fact_content":"曾在河边小学读书","confidence":0.82}`}}}})
 	}))
 	defer server.Close()
 	generator, err := NewArkGenerator(server.URL+"/api/v3", "test-key", "doubao-test", time.Second)
@@ -33,6 +33,51 @@ func TestArkGeneratorProducesValidatedContent(t *testing.T) {
 	}
 }
 
+func TestCompatibleGeneratorDiscardsSmallTalk(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		_ = json.NewEncoder(w).Encode(map[string]any{"choices": []any{map[string]any{"message": map[string]any{"content": `{"meaningful":false,"story_worthy":false,"decision_reason":"只有问候和确认","summary":"","keywords":[],"open_topic":"","story_title":"","story_content":"","fact_content":"","confidence":0.96}`}}}})
+	}))
+	defer server.Close()
+	generator, err := NewCompatibleGenerator("openai_compatible", server.URL+"/v1", "local-key", "qwen-test", 0.1, 512, time.Second)
+	if err != nil {
+		t.Fatal(err)
+	}
+	result, err := generator.Generate(context.Background(), GenerationInput{Content: "豆包：您好。\n长者：嗯，你好。\n豆包：今天怎么样？\n长者：还行。"})
+	if err != nil || result.Meaningful || result.StoryWorthy || result.AllowStory || result.DecisionReason == "" {
+		t.Fatalf("result=%+v err=%v", result, err)
+	}
+}
+
+func TestCompatibleGeneratorKeepsMeaningfulNonStoryWithoutDraft(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		_ = json.NewEncoder(w).Encode(map[string]any{"choices": []any{map[string]any{"message": map[string]any{"content": `{"meaningful":true,"story_worthy":false,"decision_reason":"有明确近况但没有事件脉络","summary":"长者今天胃口不错","keywords":["胃口","近况"],"open_topic":"下次可以聊聊最近喜欢的饭菜","story_title":"","story_content":"","fact_content":"","confidence":0.87}`}}}})
+	}))
+	defer server.Close()
+	generator, err := NewCompatibleGenerator("openai_compatible", server.URL+"/v1", "local-key", "qwen-test", 0.1, 512, time.Second)
+	if err != nil {
+		t.Fatal(err)
+	}
+	result, err := generator.Generate(context.Background(), GenerationInput{Content: "长者：我今天胃口挺好的，中午吃了面条。"})
+	if err != nil || !result.Meaningful || result.StoryWorthy || result.AllowStory || result.Summary == "" {
+		t.Fatalf("result=%+v err=%v", result, err)
+	}
+}
+
+func TestCompatibleGeneratorClearsContradictoryStoryFields(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		_ = json.NewEncoder(w).Encode(map[string]any{"choices": []any{map[string]any{"message": map[string]any{"content": `{"meaningful":true,"story_worthy":false,"decision_reason":"只是近况","summary":"今天胃口不错","keywords":["胃口"],"open_topic":"最近喜欢吃什么？","story_title":"模型误填标题","story_content":"模型误填故事","fact_content":"模型误填事实","confidence":0.8}`}}}})
+	}))
+	defer server.Close()
+	generator, err := NewCompatibleGenerator("openai_compatible", server.URL+"/v1", "local-key", "qwen-test", 0.1, 512, time.Second)
+	if err != nil {
+		t.Fatal(err)
+	}
+	result, err := generator.Generate(context.Background(), GenerationInput{Content: "长者：今天胃口不错。"})
+	if err != nil || result.AllowStory || result.StoryTitle != "" || result.StoryContent != "" || result.FactContent != "" {
+		t.Fatalf("result=%+v err=%v", result, err)
+	}
+}
+
 func TestArkGeneratorRetriesTemporaryFailure(t *testing.T) {
 	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
 		http.Error(w, "busy", http.StatusTooManyRequests)

internal/extraction/processor.go

@@ -93,7 +93,7 @@ func (p *Processor) Handle(ctx context.Context, event outbox.Event) error {
 	if len(sourceTurns) == 0 {
 		return outbox.Retryable(ErrConversationContentNotReady)
 	}
-	content := joinedTurnContent(sourceTurns)
+	content := joinedConversationContent(snapshot.turns)
 	generated, err := p.generateSafely(ctx, content, snapshot.tenantID)
 	if err != nil {
 		return err
@@ -247,12 +247,16 @@ func subjectTurns(turns []turn) []turn {
 	return items
 }
 
-func joinedTurnContent(turns []turn) string {
+func joinedConversationContent(turns []turn) string {
 	parts := make([]string, 0, len(turns))
 	for _, item := range turns {
-		parts = append(parts, item.text)
+		label := "豆包"
+		if item.speaker == "subject" {
+			label = "长者"
+		}
+		parts = append(parts, label+"："+item.text)
 	}
-	return strings.Join(parts, " ")
+	return strings.Join(parts, "\n")
 }
 
 func storeGeneration(ctx context.Context, tx pgx.Tx, conversationID string, snapshot extractionSnapshot, generated GenerationResult, generationVersion int64) error {
@@ -262,6 +266,15 @@ func storeGeneration(ctx context.Context, tx pgx.Tx, conversationID string, snap
 		sourcePublicIDs = append(sourcePublicIDs, item.publicID)
 	}
 	sourceJSON, _ := json.Marshal(sourcePublicIDs)
+	if _, err := tx.Exec(ctx, `
+		INSERT INTO conversation_content_assessment (tenant_id,subject_id,conversation_id,generation_version,meaningful,story_worthy,decision_reason,model_ref,prompt_version,confidence)
+		VALUES ($1::uuid,$2::uuid,$3::uuid,$4,$5,$6,$7,$8,$9,$10)
+		ON CONFLICT (conversation_id,generation_version) DO NOTHING`, snapshot.tenantID, snapshot.subjectID, conversationID, generationVersion, generated.Meaningful, generated.StoryWorthy, generated.DecisionReason, generated.ModelRef, generated.PromptVersion, generated.Confidence); err != nil {
+		return fmt.Errorf("store conversation content assessment: %w", err)
+	}
+	if !generated.Meaningful {
+		return tx.Commit(ctx)
+	}
 	if _, err := tx.Exec(ctx, `
 		INSERT INTO conversation_summary (tenant_id,subject_id,conversation_id,summary_type,content,source_turn_ids,model_ref,prompt_version,generation_version,confidence,status)
 		VALUES ($1::uuid,$2::uuid,$3::uuid,'conversation_summary',$4,$5::jsonb,$6,$7,$8,$9,'READY')

internal/extraction/processor_integration_test.go

@@ -40,6 +40,7 @@ func TestProcessorExtractionIdempotencyAndSensitiveContent(t *testing.T) {
 	contextID, contextPublicID := uuid.NewString(), uuid.NewString()
 	conversationID, conversationPublicID := uuid.NewString(), uuid.NewString()
 	defer func() {
+		_, _ = pool.Exec(ctx, `DELETE FROM conversation_content_assessment WHERE conversation_id=$1::uuid`, conversationID)
 		_, _ = pool.Exec(ctx, `DELETE FROM fact_candidate WHERE story_id IN (SELECT id FROM story_draft WHERE conversation_id=$1::uuid)`, conversationID)
 		_, _ = pool.Exec(ctx, `DELETE FROM story_source WHERE story_revision_id IN (SELECT sr.id FROM story_revision sr JOIN story_draft sd ON sd.id=sr.story_id WHERE sd.conversation_id=$1::uuid)`, conversationID)
 		_, _ = pool.Exec(ctx, `UPDATE story_draft SET current_revision_id=NULL WHERE conversation_id=$1::uuid`, conversationID)
@@ -97,7 +98,7 @@ func TestProcessorExtractionIdempotencyAndSensitiveContent(t *testing.T) {
 			http.Error(w, "database lock", http.StatusServiceUnavailable)
 			return
 		}
-		_, _ = io.WriteString(w, `{"choices":[{"message":{"content":"{\"summary\":\"长者谈到在河边小学教书\",\"keywords\":[\"河边小学\",\"教书经历\"],\"open_topic\":\"下次可以聊聊当年的学生\",\"story_title\":\"河边小学的回忆\",\"story_content\":\"我年轻时在河边小学教书。\",\"fact_content\":\"年轻时曾在河边小学教书\",\"confidence\":0.88}"}}]}`)
+		_, _ = io.WriteString(w, `{"choices":[{"message":{"content":"{\"meaningful\":true,\"story_worthy\":true,\"decision_reason\":\"包含具体的教书经历\",\"summary\":\"长者谈到在河边小学教书\",\"keywords\":[\"河边小学\",\"教书经历\"],\"open_topic\":\"下次可以聊聊当年的学生\",\"story_title\":\"河边小学的回忆\",\"story_content\":\"我年轻时在河边小学教书。\",\"fact_content\":\"年轻时曾在河边小学教书\",\"confidence\":0.88}"}}]}`)
 	}))
 	defer server.Close()
 	arkGenerator, err := NewArkGenerator(server.URL+"/api/v3", "integration-key", "doubao-integration", 2*time.Second)

internal/extraction/processor_safety_test.go

@@ -12,6 +12,7 @@ type safetyTestGenerator struct{ calls int }
 func (g *safetyTestGenerator) Generate(context.Context, GenerationInput) (GenerationResult, error) {
 	g.calls++
 	return GenerationResult{
+		Meaningful: true, StoryWorthy: true, DecisionReason: "测试故事",
 		Summary: "模型摘要", OpenTopic: "开放话题", StoryTitle: "标题",
 		StoryContent: "模型输出风险", FactContent: "普通事实", AllowStory: true,
 	}, nil

internal/mockprovider/server.go

@@ -391,7 +391,7 @@ func (s *Server) sendFinalTranscript(input voiceprovider.StartRequest) {
 		TaskRef: input.TaskRef, ConversationID: input.ConversationID,
 		EventType: "voice_chat.transcript.final", Status: "FINAL", OccurredAt: s.now().UTC(),
 		Turns: []voiceprovider.FinalTurn{
-			{TurnKey: "turn-1", SequenceNo: 1, Speaker: "subject", StartAt: base, EndAt: base.Add(6 * time.Second), FinalText: "我年轻时常在家乡河边散步，春天的柳树特别好看。", ProviderRef: "mock-turn-1"},
+			{TurnKey: "turn-1", SequenceNo: 1, Speaker: "subject", StartAt: base, EndAt: base.Add(6 * time.Second), FinalText: "我年轻时常和姐姐在家乡河边散步。有一次春天突然下雨，我们躲在老柳树下，后来一路笑着跑回家，现在想起来还是很高兴。", ProviderRef: "mock-turn-1"},
 			{TurnKey: "turn-2", SequenceNo: 2, Speaker: "agent", StartAt: base.Add(7 * time.Second), EndAt: base.Add(13 * time.Second), FinalText: "听起来那是一段很温暖的回忆，下次我们可以继续聊聊家乡的春天。", ProviderRef: "mock-turn-2"},
 		},
 	}

internal/privacy/delete_processor.go

@@ -266,6 +266,7 @@ func (p *DeleteProcessor) redactPostgreSQL(ctx context.Context, request deletion
 			`UPDATE subject_fact SET content='[已按隐私请求删除]',source_ref=NULL,confirmation_status='superseded',valid_to=COALESCE(valid_to,now()),version=version+1,updated_at=now() WHERE tenant_id=$1::uuid AND subject_id=$2::uuid`,
 			`UPDATE conversation_turn SET final_text='[已按隐私请求删除]',provider_ref=NULL WHERE tenant_id=$1::uuid AND subject_id=$2::uuid`,
 			`UPDATE conversation_summary SET content='[已按隐私请求删除]',source_turn_ids='["privacy-redacted"]'::jsonb,status='SUPERSEDED',last_error=NULL WHERE tenant_id=$1::uuid AND subject_id=$2::uuid`,
+			`UPDATE conversation_content_assessment SET decision_reason='[privacy-redacted]',model_ref='privacy-redacted',prompt_version='privacy-redacted' WHERE tenant_id=$1::uuid AND subject_id=$2::uuid`,
 			`UPDATE story_revision sr SET content='[已按隐私请求删除]',edit_reason=NULL,model_ref=NULL,prompt_version=NULL FROM story_draft sd WHERE sr.story_id=sd.id AND sd.tenant_id=$1::uuid AND sd.subject_id=$2::uuid`,
 			`UPDATE story_draft SET title='[已按隐私请求删除]',status='SUPERSEDED',review_reason=NULL,version=version+1,updated_at=now() WHERE tenant_id=$1::uuid AND subject_id=$2::uuid`,
 			`UPDATE fact_candidate SET content='[已按隐私请求删除]',time_range='{}'::jsonb,source_turn_ids='["privacy-redacted"]'::jsonb,status='SUPERSEDED',review_reason=NULL,version=version+1,updated_at=now() WHERE tenant_id=$1::uuid AND subject_id=$2::uuid`,

migrations/000044_conversation_content_assessment.down.sql

@@ -0,0 +1,5 @@
+BEGIN;
+
+DROP TABLE IF EXISTS conversation_content_assessment;
+
+COMMIT;

migrations/000044_conversation_content_assessment.up.sql

@@ -0,0 +1,28 @@
+BEGIN;
+
+CREATE TABLE conversation_content_assessment (
+    id                  uuid PRIMARY KEY DEFAULT gen_random_uuid(),
+    public_id           uuid NOT NULL UNIQUE DEFAULT gen_random_uuid(),
+    tenant_id           uuid NOT NULL REFERENCES tenant(id),
+    subject_id          uuid NOT NULL,
+    conversation_id     uuid NOT NULL,
+    generation_version  bigint NOT NULL CHECK (generation_version > 0),
+    meaningful          boolean NOT NULL,
+    story_worthy        boolean NOT NULL,
+    decision_reason     varchar(500) NOT NULL CHECK (length(decision_reason) > 0),
+    model_ref           varchar(160) NOT NULL,
+    prompt_version      varchar(80) NOT NULL,
+    confidence          numeric(5,4) NOT NULL CHECK (confidence BETWEEN 0 AND 1),
+    created_at          timestamptz NOT NULL DEFAULT now(),
+    CONSTRAINT conversation_content_assessment_decision_check CHECK (meaningful OR NOT story_worthy),
+    CONSTRAINT conversation_content_assessment_conversation_tenant_fk FOREIGN KEY (conversation_id, tenant_id)
+        REFERENCES conversation (id, tenant_id),
+    CONSTRAINT conversation_content_assessment_subject_tenant_fk FOREIGN KEY (subject_id, tenant_id)
+        REFERENCES subject_profile (id, tenant_id),
+    CONSTRAINT conversation_content_assessment_generation_key UNIQUE (conversation_id, generation_version)
+);
+
+CREATE INDEX conversation_content_assessment_conversation_idx
+    ON conversation_content_assessment (conversation_id, generation_version DESC);
+
+COMMIT;

scripts/smoke-conversation.ps1

@@ -98,6 +98,7 @@ DELETE FROM story_source WHERE story_revision_id IN (SELECT id FROM story_revisi
 UPDATE story_draft SET current_revision_id=NULL WHERE id IN (SELECT id FROM smoke_story_cleanup);
 DELETE FROM story_revision WHERE story_id IN (SELECT id FROM smoke_story_cleanup);
 DELETE FROM story_draft WHERE id IN (SELECT id FROM smoke_story_cleanup);
+DELETE FROM conversation_content_assessment WHERE conversation_id IN (SELECT id FROM smoke_cleanup);
 DELETE FROM conversation_summary WHERE conversation_id IN (SELECT id FROM smoke_cleanup);
 DELETE FROM conversation_turn WHERE conversation_id IN (SELECT id FROM smoke_cleanup);
 DELETE FROM provider_event WHERE conversation_id IN (SELECT id FROM smoke_cleanup);