fix: align temporary chat tools files and RAG behavior

backend/docs/docs.go

@@ -14225,9 +14225,10 @@ const docTemplate = `{
                         "BearerAuth": []
                     }
                 ],
-                "description": "由浏览器提交完整纯文本上下文；服务端不创建会话、消息、运行或断线续传记录",
+                "description": "由浏览器提交完整上下文和可选请求级附件；服务端不创建会话、消息、运行、文件或断线续传记录",
                 "consumes": [
-                    "application/json"
+                    "application/json",
+                    "multipart/form-data"
                 ],
                 "produces": [
                     "application/x-ndjson"

backend/docs/swagger.json

@@ -14218,9 +14218,10 @@
                         "BearerAuth": []
                     }
                 ],
-                "description": "由浏览器提交完整纯文本上下文；服务端不创建会话、消息、运行或断线续传记录",
+                "description": "由浏览器提交完整上下文和可选请求级附件；服务端不创建会话、消息、运行、文件或断线续传记录",
                 "consumes": [
-                    "application/json"
+                    "application/json",
+                    "multipart/form-data"
                 ],
                 "produces": [
                     "application/x-ndjson"

backend/docs/swagger.yaml

@@ -18729,7 +18729,8 @@ paths:
     post:
       consumes:
       - application/json
-      description: 由浏览器提交完整纯文本上下文；服务端不创建会话、消息、运行或断线续传记录
+      - multipart/form-data
+      description: 由浏览器提交完整上下文和可选请求级附件；服务端不创建会话、消息、运行、文件或断线续传记录
       parameters:
       - description: 临时对话参数
         in: body

backend/internal/application/contentmoderation/coordinator.go

@@ -154,12 +154,57 @@ func (c *RunCoordinator) EnqueueInputImages(ctx context.Context, fileIDs []strin
 	if len(raw) == 0 {
 		return
 	}
+	c.enqueueInputImageSources(raw, keptFiles, selected)
+}
+
+// EnqueueInputImageSources queues request-scoped image bytes without requiring
+// persisted file records. Ephemeral chat uses this path so its images follow the
+// same moderation policy while remaining outside the user file library.
+func (c *RunCoordinator) EnqueueInputImageSources(images []OutputImageSource) {
+	if c == nil {
+		return
+	}
+	selected := c.cfg.Policy.CategoriesFor(domaincm.DirectionInput, domaincm.ModalityImage)
+	if len(selected) == 0 || len(images) == 0 {
+		return
+	}
+	seenSHA := make(map[string]struct{})
+	raw := make([]OutputImageSource, 0, len(images))
+	fileIDs := make([]string, 0, len(images))
+	for _, image := range images {
+		if len(image.Data) == 0 {
+			c.recordSurfaceFailure(domaincm.DirectionInput, domaincm.ModalityImage, image.FileID, ErrModerationInvalidResp)
+			continue
+		}
+		sha := strings.TrimSpace(image.SHA256)
+		if sha != "" {
+			if _, exists := seenSHA[sha]; exists {
+				continue
+			}
+			seenSHA[sha] = struct{}{}
+		}
+		fileID := strings.TrimSpace(image.FileID)
+		raw = append(raw, OutputImageSource{
+			FileID:   fileID,
+			Data:     append([]byte(nil), image.Data...),
+			MimeType: strings.TrimSpace(image.MimeType),
+			SHA256:   sha,
+		})
+		fileIDs = append(fileIDs, fileID)
+	}
+	if len(raw) == 0 {
+		return
+	}
+	c.enqueueInputImageSources(raw, fileIDs, selected)
+}
+
+func (c *RunCoordinator) enqueueInputImageSources(raw []OutputImageSource, fileIDs []string, selected []string) {
 	c.startTask(&moderationTask{
 		Coord:     c,
 		Direction: domaincm.DirectionInput,
 		Modality:  domaincm.ModalityImage,
 		RawImages: raw,
-		FileIDs:   keptFiles,
+		FileIDs:   fileIDs,
 		Selected:  selected,
 		Location:  domaincm.ContentLocation{Field: "user_attachments"},
 		// Input hits must isolate but not revoke user library files.

backend/internal/application/contentmoderation/coordinator_test.go

@@ -343,6 +343,29 @@ func TestInputImageModerationSkipsNonImageAttachments(t *testing.T) {
 	}
 }
 
+func TestEphemeralInputImageModerationQueuesRequestScopedBytes(t *testing.T) {
+	service := NewService(nil, &coordinatorTestRepo{}, "", nil)
+	cfg := runtimeConfig{Policy: Policy{InputImageCategories: []string{"violence"}}}
+	coord := newRunCoordinator(service, RunMeta{RunID: "run_ephemeral_image", Ephemeral: true}, cfg)
+
+	coord.EnqueueInputImageSources([]OutputImageSource{
+		{FileID: "temporary_image", Data: []byte("image-bytes"), MimeType: "image/png", SHA256: "sha"},
+		{FileID: "temporary_duplicate", Data: []byte("duplicate"), MimeType: "image/png", SHA256: "sha"},
+	})
+
+	select {
+	case task := <-service.taskQueue:
+		if task == nil || task.Modality != domaincm.ModalityImage || !task.IsolateOnly {
+			t.Fatalf("unexpected request-scoped moderation task: %#v", task)
+		}
+		if len(task.RawImages) != 1 || string(task.RawImages[0].Data) != "image-bytes" {
+			t.Fatalf("request-scoped image bytes were not queued safely: %#v", task.RawImages)
+		}
+	default:
+		t.Fatal("request-scoped image moderation task was not queued")
+	}
+}
+
 func TestRecordHitRollsBackIsolatedImagesWhenEventCreateFails(t *testing.T) {
 	repo := &coordinatorTestRepo{createErr: errors.New("database unavailable")}
 	store := &coordinatorTestObjectStore{}

backend/internal/application/contentmoderation/service.go

@@ -44,7 +44,7 @@ type PreparedImage struct {
 // ImageLoader loads and prepares an image for moderation.
 type ImageLoader func(ctx context.Context, userID uint, fileID string) (PreparedImage, error)
 
-// OutputImageSource provides final generated image bytes for isolation.
+// OutputImageSource provides request-scoped image bytes and optional source metadata for moderation.
 type OutputImageSource struct {
 	FileID   string
 	Data     []byte

backend/internal/application/conversation/service_temporary_chat.go

@@ -14,26 +14,28 @@ import (
 
 const temporaryChatMaxContentChars = 1_000_000
 
-// TemporaryChatMessage 是浏览器内临时对话的一条纯文本消息。
+// TemporaryChatMessage 是浏览器内临时对话的一条消息。
 type TemporaryChatMessage struct {
 	Role    string
 	Content string
 }
 
 // TemporaryChatInput 描述不创建会话、消息和运行记录的临时推理请求。
 type TemporaryChatInput struct {
-	UserID                  uint
-	RequestID               string
-	SessionID               string
-	ClientRunID             string
-	Model                   string
-	Options                 map[string]interface{}
-	SelectedToolIDs         []uint
-	SkillIDs                []uint
-	KnowledgeBaseIDs        []string
-	HTMLVisualPromptEnabled bool
-	Messages                []TemporaryChatMessage
-	OnEvent                 func(eventType string, payload map[string]interface{}) error
+	UserID                   uint
+	RequestID                string
+	SessionID                string
+	ClientRunID              string
+	Model                    string
+	Options                  map[string]interface{}
+	SelectedToolIDs          []uint
+	SkillIDs                 []uint
+	KnowledgeBaseIDs         []string
+	HTMLVisualPromptEnabled  bool
+	Messages                 []TemporaryChatMessage
+	Attachments              []TemporaryChatAttachment
+	ReleaseAttachmentSources func()
+	OnEvent                  func(eventType string, payload map[string]interface{}) error
 }
 
 // StreamTemporaryChat 直接以请求上下文调用上游。调用方断开连接时生成随即取消，
@@ -82,6 +84,14 @@ func (s *Service) StreamTemporaryChat(
 	for _, item := range input.Messages {
 		messages = append(messages, llm.Message{Role: item.Role, Content: item.Content})
 	}
+	attachmentContext, err := s.prepareTemporaryAttachmentContext(ctx, input, messages)
+	if input.ReleaseAttachmentSources != nil {
+		input.ReleaseAttachmentSources()
+	}
+	if err != nil {
+		return nil, err
+	}
+	messages = attachmentContext.messages
 	systemPrompt := resolveMessageSystemPromptInjection(cfg, route, "", input.HTMLVisualPromptEnabled)
 	if systemPrompt.Content != "" {
 		if systemPrompt.InlineToUser {
@@ -97,8 +107,8 @@ func (s *Service) StreamTemporaryChat(
 	if err != nil {
 		return nil, err
 	}
-	// Attachment processors require persisted file objects and are therefore not
-	// exposed in a text-only temporary request.
+	// Attachment processors depend on persisted file IDs. Request-scoped temporary
+	// attachments are injected directly into the model context instead.
 	toolRuntime = toolRuntime.withoutAttachmentProcessor()
 	skillPrompts, err := s.resolveSkillPrompts(ctx, SendMessageInput{
 		UserID:   input.UserID,
@@ -115,11 +125,12 @@ func (s *Service) StreamTemporaryChat(
 		return nil, err
 	}
 	promptPlan := buildPromptPlan(ctx, promptPlanInput{
-		BaseMessages:   messages,
-		DynamicContext: knowledgeContext,
-		SkillPrompts:   skillPrompts,
-		ToolRuntime:    toolRuntime,
-		Config:         cfg,
+		BaseMessages:      messages,
+		StableAttachments: attachmentContext.stableAttachments,
+		DynamicContext:    knowledgeContext,
+		SkillPrompts:      skillPrompts,
+		ToolRuntime:       toolRuntime,
+		Config:            cfg,
 	})
 	messages = stripTemporaryMessageCacheControls(promptPlan.Messages)
 	fullMessages := cloneLLMMessages(messages)
@@ -145,15 +156,15 @@ func (s *Service) StreamTemporaryChat(
 				})
 			}
 			moderationCoord.EnqueueInputText(lastUser.Content)
+			moderationCoord.EnqueueInputImageSources(attachmentContext.moderationImages)
 		}
 	}
 	generateInput := llm.GenerateInput{
-		RequestID:    strings.TrimSpace(input.RequestID),
-		Messages:     messages,
-		Tools:        toolRuntime.definitions,
-		DisableTools: len(toolRuntime.definitions) == 0,
-		Options:      filteredOptions,
-		Ephemeral:    true,
+		RequestID: strings.TrimSpace(input.RequestID),
+		Messages:  messages,
+		Tools:     toolRuntime.definitions,
+		Options:   filteredOptions,
+		Ephemeral: true,
 	}
 	var budgetFit promptBudgetFit
 	generateInput, budgetFit = fitGenerateInputToModelBudget(
@@ -345,6 +356,9 @@ func ValidateTemporaryChatInput(input TemporaryChatInput) error {
 	if len(input.Messages) == 0 || len(input.Messages) > 100 {
 		return ErrInvalidMessageContent
 	}
+	if len(input.Attachments) > TemporaryChatMaxAttachments {
+		return ErrTooManyMessageFiles
+	}
 	if len(input.KnowledgeBaseIDs) > 8 {
 		return ErrInvalidMessageContent
 	}
@@ -359,12 +373,22 @@ func ValidateTemporaryChatInput(input TemporaryChatInput) error {
 		}
 		seenKnowledgeBases[normalized] = struct{}{}
 	}
+	attachmentCounts := make(map[int]int)
+	for _, attachment := range input.Attachments {
+		if attachment.MessageIndex < 0 || attachment.MessageIndex >= len(input.Messages) || attachment.Reader == nil {
+			return ErrInvalidFileReference
+		}
+		attachmentCounts[attachment.MessageIndex]++
+	}
 	totalChars := 0
 	previousRole := ""
-	for _, item := range input.Messages {
+	for index, item := range input.Messages {
 		role := strings.TrimSpace(item.Role)
 		content := strings.TrimSpace(item.Content)
-		if (role != "user" && role != "assistant") || content == "" || role == previousRole {
+		if (role != "user" && role != "assistant") || role == previousRole {
+			return ErrInvalidMessageContent
+		}
+		if content == "" && (role != "user" || attachmentCounts[index] == 0) {
 			return ErrInvalidMessageContent
 		}
 		totalChars += len([]rune(item.Content))