Skip to content

Open-source LLM knowledge platform: turn raw documents into a queryable RAG, an autonomous reasoning agent, and a self-maintaining Wiki.

51.3 KB
Raw Blame
package service

import (
	"context"
	"fmt"
	"strings"
	"time"

	"github.com/Tencent/WeKnora/internal/agent/tools"
	chatpipeline "github.com/Tencent/WeKnora/internal/application/service/chat_pipeline"
	"github.com/Tencent/WeKnora/internal/common"
	apperrors "github.com/Tencent/WeKnora/internal/errors"
	"github.com/Tencent/WeKnora/internal/event"
	"github.com/Tencent/WeKnora/internal/logger"
	"github.com/Tencent/WeKnora/internal/modelcontext"
	"github.com/Tencent/WeKnora/internal/models/chat"
	"github.com/Tencent/WeKnora/internal/tracing/langfuse"
	"github.com/Tencent/WeKnora/internal/types"
)

// KnowledgeQA performs knowledge base question answering with LLM summarization
// Events are emitted through eventBus (references, answer chunks, completion)
// customAgent is optional - if provided, uses custom agent configuration for multiTurnEnabled and historyTurns
func (s *sessionService) KnowledgeQA(
	ctx context.Context,
	req *types.QARequest,
	eventBus *event.EventBus,
) error {
	webSearchEnabled := resolveWebSearchEnabled(req)
	logger.Infof(
		ctx,
		"Knowledge base question answering parameters, session ID: %s, query: %s, webSearchEnabled: %v",
		req.Session.ID,
		req.Query,
		webSearchEnabled,
	)
	// IM/MCP call KnowledgeQA without executeQA's send-side lease, so hold one
	// for the pipeline. HTTP send already took it before persisting the turn
	// and says so on the request; taking a second would double every rewind
	// check and lease round trip on the hot send path.
	if !req.TurnLeaseHeld {
		if err := s.RejectSendIfRewinding(ctx, req.Session.ID); err != nil {
			return err
		}
		configID := ""
		if req.CustomAgent != nil {
			configID = req.CustomAgent.Config.SandboxConfigID
		}
		turnCtx, releaseTurn, holdErr := s.holdSandboxTurn(ctx, req.Session.ID, configID)
		if holdErr != nil {
			return holdErr
		}
		defer releaseTurn()
		ctx = turnCtx
	}

	// Span the request setup (KB / model resolution, search target building,
	// agent override application). This covers the visible gap between trace
	// start and the first stage observation in the Langfuse timeline.
	setupCtx, setupSpan := langfuse.GetManager().StartSpan(ctx, langfuse.SpanOptions{
		Name: "qa.setup",
		Metadata: map[string]interface{}{
			"session_id": req.Session.ID,
		},
	})
	ctx = setupCtx

	// Resolve knowledge bases using shared helper
	knowledgeBaseIDs, knowledgeIDs, err := s.resolveKnowledgeBases(ctx, req)
	if err != nil {
		return err
	}

	// Resolve chat model ID using shared helper
	chatModelID, chatModelTenantID, err := s.resolveChatModelID(ctx, req, knowledgeBaseIDs, knowledgeIDs)
	if err != nil {
		return err
	}

	// Initialize ChatManage defaults from config.yaml
	summaryConfig := types.SummaryConfig{
		Prompt:              s.cfg.Conversation.Summary.Prompt,
		ContextTemplate:     s.cfg.Conversation.Summary.ContextTemplate,
		Temperature:         s.cfg.Conversation.Summary.Temperature,
		NoMatchPrefix:       s.cfg.Conversation.Summary.NoMatchPrefix,
		MaxCompletionTokens: s.cfg.Conversation.Summary.MaxCompletionTokens,
		Thinking:            s.cfg.Conversation.Summary.Thinking,
	}
	fallbackStrategy := types.FallbackStrategy(s.cfg.Conversation.FallbackStrategy)
	if fallbackStrategy == "" {
		fallbackStrategy = types.FallbackStrategyFixed
		logger.Infof(ctx, "Fallback strategy not set, using default: %v", fallbackStrategy)
	}

	// Resolve chat model vision capability and VLM model ID for image routing
	var chatModelSupportsVision bool
	var vlmModelID string
	if chatModelID != "" {
		modelCtx := types.WithExecutionTenant(ctx, chatModelTenantID)
		chatModelInfo, err := s.modelService.GetModelByID(modelCtx, chatModelID)
		if err == nil && chatModelInfo != nil {
			chatModelSupportsVision = chatModelInfo.Parameters.SupportsVision
		}
	}
	if req.CustomAgent != nil {
		vlmModelID = req.CustomAgent.Config.VLMModelID
	}

	// Resolve retrieval tenant scope using shared helper
	retrievalTenantID := s.resolveRetrievalTenantID(ctx, req)

	// Build unified search targets (computed once, used throughout pipeline)
	searchTargets, err := s.buildSearchTargets(ctx, retrievalTenantID, knowledgeBaseIDs, knowledgeIDs, req.TagScopes)
	if err != nil {
		return fmt.Errorf("build search targets: %w", err)
	}

	// Create chat management object with session settings
	logger.Infof(
		ctx,
		"Creating chat manage object, knowledge base IDs: %v, knowledge IDs: %v, chat model ID: %s, search targets: %d",
		knowledgeBaseIDs,
		knowledgeIDs,
		chatModelID,
		len(searchTargets),
	)

	chatManage := &types.ChatManage{
		PipelineRequest: types.PipelineRequest{
			Query:                   req.Query,
			SessionID:               req.Session.ID,
			UserID:                  types.SessionOwnerIDFromContext(ctx),
			MaxRounds:               s.cfg.Conversation.MaxRounds,
			KnowledgeBaseIDs:        knowledgeBaseIDs,
			KnowledgeIDs:            knowledgeIDs,
			SearchTargets:           searchTargets,
			VectorThreshold:         s.cfg.Conversation.VectorThreshold,
			KeywordThreshold:        s.cfg.Conversation.KeywordThreshold,
			EmbeddingTopK:           s.cfg.Conversation.EmbeddingTopK,
			RerankTopK:              s.cfg.Conversation.RerankTopK,
			RerankThreshold:         s.cfg.Conversation.RerankThreshold,
			ChatModelID:             chatModelID,
			ChatModelTenantID:       chatModelTenantID,
			SummaryConfig:           summaryConfig,
			FallbackStrategy:        fallbackStrategy,
			FallbackResponse:        s.cfg.Conversation.FallbackResponse,
			FallbackPrompt:          s.cfg.Conversation.FallbackPrompt,
			EnableRewrite:           s.cfg.Conversation.EnableRewrite,
			EnableQueryExpansion:    s.cfg.Conversation.EnableQueryExpansion,
			RewritePromptSystem:     s.cfg.Conversation.RewritePromptSystem,
			RewritePromptUser:       s.cfg.Conversation.RewritePromptUser,
			WebSearchEnabled:        webSearchEnabled,
			WebSearchProviderID:     s.resolveWebSearchProviderID(ctx, req, retrievalTenantID),
			WebSearchMaxResults:     s.resolveWebSearchMaxResults(ctx, req),
			WebFetchEnabled:         s.resolveWebFetchEnabled(req),
			WebFetchTopN:            s.resolveWebFetchTopN(req),
			TenantID:                retrievalTenantID,
			Images:                  req.ImageURLs,
			VLMModelID:              vlmModelID,
			ChatModelSupportsVision: chatModelSupportsVision,
			Attachments:             req.Attachments,
			Language:                types.LanguageNameFromContext(ctx),
		},
		PipelineState: types.PipelineState{
			RewriteQuery:     req.Query,
			ImageDescription: req.ImageDescription,
			QuotedContext:    req.QuotedContext,
		},
		PipelineContext: types.PipelineContext{
			EventBus:      eventBus.AsEventBusInterface(),
			MessageID:     req.AssistantMessageID,
			UserMessageID: req.UserMessageID,
		},
	}

	// Apply custom agent overrides (system prompt, temperature, retrieval params,
	// rewrite, fallback, FAQ strategy, history turns)
	s.applyAgentOverridesToChatManage(ctx, req.CustomAgent, chatManage)
	applyRequestReasoningEffort(
		req.ReasoningEffort, &chatManage.SummaryConfig.Thinking, &chatManage.SummaryConfig.ReasoningEffort,
	)

	// An agent may opt out of long-term memory. The preference is per-request
	// rather than per-user, so it travels in the context that the recall
	// plugin reads.
	if req.CustomAgent != nil {
		ctx = types.ApplyAgentMemoryPreference(ctx, req.CustomAgent.Config.MemoryEnabled)
	}

	// Determine pipeline based on the effective knowledge retrieval scope and
	// web search setting. Tag-only mentions leave the raw KB/knowledge ID slices
	// empty but produce SearchTargets, so the unified targets must participate in
	// this decision or the request is incorrectly downgraded to pure chat.
	hasKB := types.HasKnowledgeRetrievalScope(searchTargets, knowledgeBaseIDs, knowledgeIDs)
	needsRAG := hasKB || webSearchEnabled
	hasHistory := chatManage.MaxRounds > 0

	var pipeline []types.EventType
	if !needsRAG {
		// Pure chat — no retrieval needed.
		userContent := req.Query
		if req.ImageDescription != "" && !chatModelSupportsVision {
			userContent += "\n\n[用户上传图片内容]\n" + req.ImageDescription
		}
		if req.QuotedContext != "" {
			userContent += "\n\n" + req.QuotedContext
		}
		// Inject attachment content for pure-chat path (RAG path handles this in INTO_CHAT_MESSAGE).
		if len(req.Attachments) > 0 {
			userContent += req.Attachments.BuildPrompt()
		}
		chatManage.UserContent = userContent

		pipeline = types.NewPipelineBuilder().
			AddIf(hasHistory, types.LOAD_HISTORY).
			Add(types.MEMORY_RECALL).
			Add(types.CHAT_COMPLETION_STREAM).
			Build()
	} else {
		// RAG — dynamically assemble based on feature flags.
		pipeline = types.NewPipelineBuilder().
			AddIf(hasHistory, types.LOAD_HISTORY).
			Add(types.MEMORY_RECALL).
			Add(types.QUERY_UNDERSTAND).
			Add(types.CHUNK_SEARCH_PARALLEL).
			Add(types.CHUNK_RERANK).
			AddIf(webSearchEnabled, types.WEB_FETCH).
			Add(types.CHUNK_MERGE).
			Add(types.FILTER_TOP_K).
			AddIf(chatManage.DataAnalysisEnabled, types.DATA_ANALYSIS).
			Add(types.INTO_CHAT_MESSAGE).
			Add(types.CHAT_COMPLETION_STREAM).
			Build()
	}

	logger.Infof(ctx, "Assembled pipeline (%d stages), hasKB=%v, webSearch=%v, history=%v",
		len(pipeline), hasKB, webSearchEnabled, hasHistory)

	// Start knowledge QA event processing (set session tenant so pipeline session/message lookups use session owner)
	ctx = context.WithValue(ctx, types.SessionTenantIDContextKey, req.Session.TenantID)
	// Propagate the session ID so stateful sandbox backends (CubeSandbox) can
	// bind script execution to a per-session MicroVM instance.
	ctx = types.WithSessionID(ctx, req.Session.ID)
	logger.Info(ctx, "Triggering question answering event")
	setupSpan.Finish(map[string]interface{}{
		"stages":             len(pipeline),
		"knowledge_base_ids": knowledgeBaseIDs,
		"search_targets":     len(searchTargets),
	}, nil, nil)
	err = s.KnowledgeQAByEvent(ctx, chatManage, pipeline)
	if err != nil {
		logger.ErrorWithFields(ctx, err, map[string]interface{}{
			"session_id": req.Session.ID,
		})
		return err
	}

	// Note: Answer events are now emitted directly by chat_completion_stream plugin
	// Completion event will be emitted when the last answer event has Done=true
	// We can optionally add a completion watcher here if needed, but for now
	// the frontend can detect completion from the Done flag

	logger.Info(ctx, "Knowledge base question answering initiated")
	return nil
}

// selectChatModelID prefers a remote model from an accessible KB, then the
// first accessible KB's model, then an available model in the execution tenant.
// Return the KB owner alongside its model ID without changing the request scope.
func (s *sessionService) selectChatModelID(
	ctx context.Context,
	knowledgeBaseIDs []string,
	knowledgeIDs []string,
) (string, uint64, error) {
	tenantID, _ := types.TenantIDFromContext(ctx)
	if len(knowledgeBaseIDs) == 0 && len(knowledgeIDs) > 0 {
		knowledgeList, err := s.knowledgeService.GetKnowledgeBatchWithSharedAccess(ctx, tenantID, knowledgeIDs)
		if err != nil {
			return "", 0, fmt.Errorf("resolve knowledge bases for chat model: %w", err)
		}
		for _, k := range knowledgeList {
			if k != nil && k.KnowledgeBaseID != "" {
				knowledgeBaseIDs = append(knowledgeBaseIDs, k.KnowledgeBaseID)
			}
		}
		knowledgeBaseIDs = uniqueNonEmptyStrings(knowledgeBaseIDs)
	}

	permissions := kbReadPermissions(ctx, s.kbShareService)
	var firstKB *types.KnowledgeBase
	for _, kbID := range knowledgeBaseIDs {
		kb, err := s.knowledgeBaseService.GetKnowledgeBaseByID(ctx, kbID)
		if err != nil || kb == nil {
			logger.Warnf(ctx, "Failed to get knowledge base %s for chat model: %v", kbID, err)
			continue
		}
		allowed, err := permissions.Check(kb.ID, kb.TenantID, types.OrgRoleViewer)
		if err != nil || !allowed {
			logger.Warnf(ctx, "Cannot use chat model from inaccessible knowledge base %s", kb.ID)
			continue
		}
		if firstKB == nil {
			firstKB = kb
		}
		if kb.SummaryModelID == "" {
			continue
		}
		modelCtx := types.WithExecutionTenant(ctx, kb.TenantID)
		model, err := s.modelService.GetModelByID(modelCtx, kb.SummaryModelID)
		if err == nil && model != nil && model.Source == types.ModelSourceRemote {
			return kb.SummaryModelID, kb.TenantID, nil
		}
	}
	if firstKB != nil && firstKB.SummaryModelID != "" {
		return firstKB.SummaryModelID, firstKB.TenantID, nil
	}

	models, err := s.modelService.ListModels(ctx)
	if err != nil {
		return "", 0, fmt.Errorf("failed to list models: %w", err)
	}
	if model := types.PickChatModel(models, ""); model != nil {
		return model.ID, tenantID, nil
	}
	return "", 0, fmt.Errorf("no chat model ID available: no accessible knowledge base model or available models")
}

// resolveKnowledgeBasesFromAgent resolves knowledge base IDs based on agent's KBSelectionMode.
// sessionTenantID is the tenant of the current session (caller); it is compared with
// customAgent.TenantID to detect the shared-agent scenario and avoid leaking the
// current user's personal shared KBs into the agent's retrieval scope.
//
// Returns the resolved knowledge base IDs based on the selection mode:
//   - "all": fetches all knowledge bases for the tenant
//   - "selected": uses the explicitly configured knowledge bases
//   - "none": returns empty slice
//   - default: falls back to configured knowledge bases for backward compatibility
func (s *sessionService) resolveKnowledgeBasesFromAgent(
	ctx context.Context,
	customAgent *types.CustomAgent,
	sessionTenantID uint64,
) []string {
	if customAgent == nil {
		return nil
	}

	switch customAgent.Config.KBSelectionMode {
	case "all":
		// Authoritative capability filter for the runtime path. The frontend
		// editor and @mention dropdown apply the same filter, but we don't
		// trust the client here: a stale session payload or API caller could
		// still ask us to retrieve against an incompatible KB and we'd rather
		// just drop it (and log) than feed it to tools that would no-op.
		capFilter := tools.DeriveKBFilterForAgent(customAgent.Config.AgentMode, customAgent.Config.AllowedTools)
		accept := func(kb *types.KnowledgeBase) bool {
			if kb == nil {
				return false
			}
			if capFilter.IsEmpty() {
				return true
			}
			return tools.KBSatisfiesAgentRequirements(kb.Capabilities(), customAgent.Config.AgentMode, customAgent.Config.AllowedTools)
		}

		// Get own knowledge bases (uses ctx TenantID = agent's tenant)
		allKBs, err := s.knowledgeBaseService.ListKnowledgeBases(ctx)
		if err != nil {
			logger.Warnf(ctx, "Failed to list all knowledge bases: %v", err)
		}
		kbIDSet := make(map[string]bool)
		kbIDs := make([]string, 0, len(allKBs))
		ownSkipped := 0
		for _, kb := range allKBs {
			if !accept(kb) {
				ownSkipped++
				continue
			}
			kbIDs = append(kbIDs, kb.ID)
			kbIDSet[kb.ID] = true
		}

		// For shared agents (session tenant != agent tenant), only use the agent
		// tenant's own KBs. Including the current user's shared KBs would leak
		// unrelated KBs from other organisations into the agent's retrieval scope.
		isSharedAgent := sessionTenantID != 0 && sessionTenantID != customAgent.TenantID
		sharedSkipped := 0
		if !isSharedAgent {
			tenantID := types.MustTenantIDFromContext(ctx)
			userIDVal := ctx.Value(types.UserIDContextKey)
			if userIDVal != nil {
				if userID, ok := userIDVal.(string); ok && userID != "" && s.kbShareService != nil {
					callerTenantRole := types.TenantRoleFromContext(ctx)
					sharedList, err := s.kbShareService.ListSharedKnowledgeBases(ctx, tenantID, callerTenantRole)
					if err != nil {
						logger.Warnf(ctx, "Failed to list shared knowledge bases: %v", err)
					} else {
						for _, info := range sharedList {
							if info == nil || info.KnowledgeBase == nil || kbIDSet[info.KnowledgeBase.ID] {
								continue
							}
							if !accept(info.KnowledgeBase) {
								sharedSkipped++
								continue
							}
							kbIDs = append(kbIDs, info.KnowledgeBase.ID)
							kbIDSet[info.KnowledgeBase.ID] = true
						}
					}
				}
			}
		} else {
			logger.Infof(ctx, "Shared agent detected (session tenant %d != agent tenant %d): skipping user's shared KBs",
				sessionTenantID, customAgent.TenantID)
		}

		if ownSkipped+sharedSkipped > 0 {
			logger.Infof(ctx,
				"KBSelectionMode=all: tool-capability filter removed %d own + %d shared KBs (agent=%s, tools=%v)",
				ownSkipped, sharedSkipped, customAgent.ID, customAgent.Config.AllowedTools)
		}
		logger.Infof(ctx, "KBSelectionMode=all: loaded %d knowledge bases (own + shared)", len(kbIDs))
		return kbIDs
	case "selected":
		logger.Infof(ctx, "KBSelectionMode=selected: using %d configured knowledge bases", len(customAgent.Config.KnowledgeBases))
		return customAgent.Config.KnowledgeBases
	case "none":
		logger.Infof(ctx, "KBSelectionMode=none: no knowledge bases configured")
		return nil
	default:
		// Default to "selected" behavior for backward compatibility
		if len(customAgent.Config.KnowledgeBases) > 0 {
			logger.Infof(ctx, "KBSelectionMode not set: using %d configured knowledge bases", len(customAgent.Config.KnowledgeBases))
		}
		return customAgent.Config.KnowledgeBases
	}
}

// buildSearchTargets computes the unified search targets from knowledgeBaseIDs and knowledgeIDs.
// tenantID is the retrieval scope: session.TenantID or effective tenant from shared agent (set by handler).
// This is called once at the request entry point to avoid repeated queries later in the pipeline.
// Logic:
//   - For each knowledgeBaseID: resolve actual TenantID (own, org-shared, or in retrieval-tenant scope for shared agent)
//   - For each knowledgeID: find its knowledgeBaseID; if the KB is already in the list, skip; otherwise add SearchTargetTypeKnowledge
func (s *sessionService) buildSearchTargets(
	ctx context.Context,
	tenantID uint64,
	knowledgeBaseIDs []string,
	knowledgeIDs []string,
	tagScopes []types.TagScope,
) (types.SearchTargets, error) {
	caller := types.CallerFromContext(ctx)
	var targets types.SearchTargets
	tagIDsByKB := mergeTagScopesByKB(tagScopes)

	// Build a map from KB ID to TenantID for all KBs we need to process
	kbTenantMap := make(map[string]uint64)

	// Track which KBs are fully searched
	fullKBSet := make(map[string]bool)

	// First pass: batch-fetch KBs, then resolve tenant per ID (tenant scope already set by caller)
	permissions := kbReadPermissions(ctx, s.kbShareService)
	kbIDsToFetch := append([]string(nil), knowledgeBaseIDs...)
	for kbID := range tagIDsByKB {
		kbIDsToFetch = append(kbIDsToFetch, kbID)
	}
	kbIDsToFetch = uniqueNonEmptyStrings(kbIDsToFetch)

	kbByID := make(map[string]*types.KnowledgeBase)
	if len(kbIDsToFetch) > 0 {
		kbs, kbFetchErr := s.knowledgeBaseService.GetKnowledgeBasesByIDsOnly(ctx, kbIDsToFetch)
		if kbFetchErr != nil {
			logger.Warnf(ctx, "Failed to fetch knowledge bases for search targets: %v", kbFetchErr)
		}
		for _, kb := range kbs {
			if kb != nil {
				kbByID[kb.ID] = kb
			}
		}
	}
	resolveKBTenant := func(kbID string) uint64 {
		if kbTenantMap[kbID] != 0 {
			return kbTenantMap[kbID]
		}
		kb := kbByID[kbID]
		kbTenantMap[kbID] = caller.TenantID
		if kb != nil {
			kbTenantMap[kbID] = 0
			if allowed, err := permissions.Check(kbID, kb.TenantID, types.OrgRoleViewer); err == nil && allowed {
				kbTenantMap[kbID] = kb.TenantID
			}
		}
		return kbTenantMap[kbID]
	}

	if len(knowledgeBaseIDs) > 0 {
		for _, kbID := range knowledgeBaseIDs {
			fullKBSet[kbID] = true
			kbTenant := resolveKBTenant(kbID)
			if kbTenant == 0 {
				continue
			}
			if len(tagIDsByKB[kbID]) > 0 {
				continue
			}
			targets = append(targets, &types.SearchTarget{
				Type:            types.SearchTargetTypeKnowledgeBase,
				KnowledgeBaseID: kbID,
				TenantID:        kbTenant,
			})
		}
	}

	kbToKnowledgeIDs := make(map[string][]string)

	// Process individual knowledge IDs (include shared KB files the user has access to)
	if len(knowledgeIDs) > 0 {
		knowledgeList, err := s.knowledgeService.GetKnowledgeBatchWithSharedAccess(ctx, tenantID, knowledgeIDs)
		if err != nil {
			logger.Warnf(ctx, "Failed to get knowledge batch for search targets: %v", err)
			return targets, nil // Return what we have, don't fail
		}

		// Group knowledge IDs by their KB, excluding those already covered by full KB search
		// Also track KB tenant IDs from knowledge items
		for _, k := range knowledgeList {
			if k == nil || k.KnowledgeBaseID == "" {
				continue
			}
			// Track KB -> TenantID mapping from knowledge items
			if kbTenantMap[k.KnowledgeBaseID] == 0 {
				kbTenantMap[k.KnowledgeBaseID] = k.TenantID
			}
			// Skip if this KB is already fully searched without a tag scope.
			if fullKBSet[k.KnowledgeBaseID] && len(tagIDsByKB[k.KnowledgeBaseID]) == 0 {
				continue
			}
			kbToKnowledgeIDs[k.KnowledgeBaseID] = append(kbToKnowledgeIDs[k.KnowledgeBaseID], k.ID)
		}

		// Create SearchTargetTypeKnowledge targets for each KB with specific files
		for kbID, kidList := range kbToKnowledgeIDs {
			if len(tagIDsByKB[kbID]) > 0 {
				continue
			}
			kbTenant := kbTenantMap[kbID]
			if kbTenant == 0 {
				kbTenant = tenantID // fallback
			}
			targets = append(targets, &types.SearchTarget{
				Type:                    types.SearchTargetTypeKnowledge,
				KnowledgeBaseID:         kbID,
				TenantID:                kbTenant,
				KnowledgeIDs:            kidList,
				DisableRecallThresholds: true,
			})
		}
	}

	for kbID, tagIDs := range tagIDsByKB {
		if kbID == "" || len(tagIDs) == 0 {
			continue
		}
		kbTenant := resolveKBTenant(kbID)
		if kbTenant == 0 {
			continue
		}
		kb := kbByID[kbID]
		explicitKnowledgeIDs := uniqueNonEmptyStrings(kbToKnowledgeIDs[kbID])

		useDocumentTagResolution := kb == nil || kb.Type != types.KnowledgeBaseTypeFAQ
		if kb == nil {
			logger.Warnf(ctx, "Knowledge base metadata missing for tag scope, kb_id=%s, using document tag resolution", kbID)
		}
		if useDocumentTagResolution {
			tagKnowledgeIDs, err := s.knowledgeService.ListKnowledgeIDsByTagIDs(ctx, kbTenant, kbID, tagIDs)
			if err != nil {
				return nil, fmt.Errorf("resolve knowledge IDs for tag scope kb_id=%s: %w", kbID, err)
			}
			if len(explicitKnowledgeIDs) > 0 {
				tagKnowledgeIDs = intersectStrings(tagKnowledgeIDs, explicitKnowledgeIDs)
			}
			tagKnowledgeIDs = uniqueNonEmptyStrings(tagKnowledgeIDs)
			if len(tagKnowledgeIDs) == 0 {
				continue
			}
			targets = append(targets, &types.SearchTarget{
				Type:                    types.SearchTargetTypeKnowledge,
				KnowledgeBaseID:         kbID,
				TenantID:                kbTenant,
				KnowledgeIDs:            tagKnowledgeIDs,
				ScopeTagIDs:             append([]string(nil), tagIDs...),
				DisableRecallThresholds: true,
			})
			continue
		}

		target := &types.SearchTarget{
			Type:                    types.SearchTargetTypeKnowledgeBase,
			KnowledgeBaseID:         kbID,
			TenantID:                kbTenant,
			TagIDs:                  append([]string(nil), tagIDs...),
			ScopeTagIDs:             append([]string(nil), tagIDs...),
			DisableRecallThresholds: true,
		}
		if len(explicitKnowledgeIDs) > 0 {
			target.Type = types.SearchTargetTypeKnowledge
			target.KnowledgeIDs = explicitKnowledgeIDs
			target.DisableRecallThresholds = true
		}
		targets = append(targets, target)
	}

	logger.Infof(ctx, "Built %d search targets: %d full KB, %d partial/tag KB, kbTenantMap=%v",
		len(targets), len(knowledgeBaseIDs), len(targets)-len(knowledgeBaseIDs), kbTenantMap)

	return targets, nil
}

func mergeTagScopesByKB(scopes []types.TagScope) map[string][]string {
	byKB := make(map[string][]string)
	seen := make(map[string]map[string]bool)
	for _, scope := range scopes {
		if scope.KnowledgeBaseID == "" {
			continue
		}
		if seen[scope.KnowledgeBaseID] == nil {
			seen[scope.KnowledgeBaseID] = make(map[string]bool)
		}
		for _, tagID := range scope.TagIDs {
			if tagID == "" || seen[scope.KnowledgeBaseID][tagID] {
				continue
			}
			seen[scope.KnowledgeBaseID][tagID] = true
			byKB[scope.KnowledgeBaseID] = append(byKB[scope.KnowledgeBaseID], tagID)
		}
	}
	return byKB
}

func uniqueNonEmptyStrings(values []string) []string {
	seen := make(map[string]bool, len(values))
	out := make([]string, 0, len(values))
	for _, value := range values {
		if value == "" || seen[value] {
			continue
		}
		seen[value] = true
		out = append(out, value)
	}
	return out
}

func intersectStrings(left []string, right []string) []string {
	if len(left) == 0 || len(right) == 0 {
		return nil
	}
	rightSet := make(map[string]bool, len(right))
	for _, value := range right {
		rightSet[value] = true
	}
	out := make([]string, 0)
	for _, value := range left {
		if rightSet[value] {
			out = append(out, value)
		}
	}
	return out
}

// KnowledgeQAByEvent processes knowledge QA through a series of events in the pipeline
func (s *sessionService) KnowledgeQAByEvent(ctx context.Context,
	chatManage *types.ChatManage, eventList []types.EventType,
) error {
	ctx = types.WithSessionID(ctx, chatManage.SessionID)
	logger.Info(ctx, "Start processing knowledge base question answering through events")
	logger.Infof(ctx, "Knowledge base question answering parameters, session ID: %s, query: %s",
		chatManage.SessionID, chatManage.Query)

	methods := make([]string, len(eventList))
	for i, event := range eventList {
		methods[i] = string(event)
	}
	logger.Infof(ctx, "Trigger event list: %v", methods)

	pipelineStart := time.Now()
	lastRetrievalStage := chatpipeline.LastConsolidatedRetrievalStage(eventList, chatManage)
	var retrievalProgress *chatpipeline.StageProgress
	var retrievalStart time.Time
	var understandProgress *chatpipeline.StageProgress
	var understandStart time.Time
	for _, eventType := range eventList {
		stageStart := time.Now()
		// Wrap each pipeline stage in a Langfuse span so the trace timeline
		// shows the gaps between LLM/embedding/rerank generations (the work
		// that happens between them — vector DB search, merge, filter, prompt
		// assembly — was previously invisible). Generations created inside
		// the stage automatically nest under this span.
		//
		// CHAT_COMPLETION_STREAM is intentionally skipped: its OnEvent kicks
		// off a streaming goroutine and returns immediately, so a span would
		// finish well before the chat.completion.stream generation does. The
		// generation already captures the full stream duration; adding a
		// stage span here would just produce a child observation that
		// visually exceeds its parent.
		stageCtx := ctx
		var stageSpan *langfuse.Span
		if eventType != types.CHAT_COMPLETION_STREAM {
			stageCtx, stageSpan = langfuse.GetManager().StartSpan(ctx, langfuse.SpanOptions{
				Name: "pipeline." + string(eventType),
				Metadata: map[string]interface{}{
					"event_type": string(eventType),
					"session_id": chatManage.SessionID,
				},
			})
		}
		if eventType == types.QUERY_UNDERSTAND && chatpipeline.ShouldEmitQueryUnderstandProgress(chatManage) {
			understandStart = stageStart
			understandProgress = chatpipeline.BeginQueryUnderstandProgress(stageCtx, chatManage)
		}
		if chatpipeline.IsConsolidatedRetrievalStage(eventType, chatManage) && retrievalProgress == nil {
			retrievalStart = stageStart
			retrievalProgress = chatpipeline.BeginRetrievalProgress(stageCtx, chatManage)
		}
		// Emit references before answer streaming so the SSE client receives
		// them while the connection is still open. Previously references were
		// emitted after the pipeline returned — by then the `complete` event had
		// already closed the stream, so the frontend only saw citations on refresh.
		if eventType == types.CHAT_COMPLETION_STREAM {
			emitKnowledgeReferencesEvent(ctx, chatManage)
		}
		err := s.eventManager.Trigger(stageCtx, eventType, chatManage)
		if understandProgress != nil && eventType == types.QUERY_UNDERSTAND {
			chatpipeline.EndQueryUnderstandProgress(stageCtx, chatManage, understandProgress, understandStart, err)
			understandProgress = nil
		}
		// Close the consolidated retrieval progress window as soon as retrieval
		// is done: either the planned last retrieval stage completed, or a
		// retrieval stage short-circuited the pipeline (ErrSearchNothing or a
		// hard error). The early returns below (fallback / stage_failed) would
		// otherwise skip EndRetrievalProgress, leaving the "knowledge_search"
		// tool_call pending — so the frontend keeps spinning on "正在检索知识库"
		// forever even though the fallback answer has already streamed.
		if retrievalProgress != nil && chatpipeline.ShouldCloseRetrievalProgress(eventType, lastRetrievalStage, err) {
			chatpipeline.EndRetrievalProgress(stageCtx, chatManage, retrievalProgress, retrievalStart, err)
			retrievalProgress = nil
		}
		stageDuration := time.Since(stageStart)
		var spanErr error
		if err != nil && err != chatpipeline.ErrSearchNothing {
			spanErr = err.Err
		}
		if stageSpan != nil {
			stageSpan.Finish(map[string]interface{}{
				"duration_ms": stageDuration.Milliseconds(),
			}, nil, spanErr)
		}

		// If the user stopped generation, the context is cancelled. A cancelled
		// retrieval stage surfaces as ErrSearchNothing (the search goroutines
		// return no results when their embedding/vector calls are aborted), so
		// this check MUST come before the ErrSearchNothing handling below.
		// Otherwise we would persist the fixed fallback response ("Sorry, I am
		// unable to answer this question.") over the intentionally-empty stopped
		// message, and the user would see the fallback text after refreshing.
		// This is not single-machine specific: the stop arrives via the shared
		// StreamManager and cancels asyncCtx on whichever node is generating.
		if ctxErr := ctx.Err(); ctxErr != nil {
			common.PipelineWarn(ctx, "Pipeline", "stage_cancelled", map[string]interface{}{
				"event":       string(eventType),
				"duration_ms": stageDuration.Milliseconds(),
				"reason":      ctxErr.Error(),
			})
			return ctxErr
		}

		if err == chatpipeline.ErrSearchNothing {
			common.PipelineWarn(ctx, "Pipeline", "stage_fallback", map[string]interface{}{
				"event":       string(eventType),
				"duration_ms": stageDuration.Milliseconds(),
				"reason":      "search_nothing",
				"strategy":    string(chatManage.FallbackStrategy),
			})
			s.handleFallbackResponse(ctx, chatManage)
			return nil
		}

		if err != nil {
			common.PipelineError(ctx, "Pipeline", "stage_failed", map[string]interface{}{
				"event":       string(eventType),
				"duration_ms": stageDuration.Milliseconds(),
				"error_type":  err.ErrorType,
				"description": err.Description,
			})
			return err.Err
		}

		common.PipelineInfo(ctx, "Pipeline", "stage_complete", map[string]interface{}{
			"event":       string(eventType),
			"duration_ms": stageDuration.Milliseconds(),
		})
	}

	common.PipelineInfo(ctx, "Pipeline", "all_stages_complete", map[string]interface{}{
		"session_id":        chatManage.SessionID,
		"total_stages":      len(eventList),
		"total_duration_ms": time.Since(pipelineStart).Milliseconds(),
	})
	return nil
}

// SearchKnowledge performs knowledge base search without LLM summarization
// knowledgeBaseIDs: list of knowledge base IDs to search (supports multi-KB)
// knowledgeIDs: list of specific knowledge (file) IDs to search
// opts: caller overrides of the tenant RetrievalConfig; nil keeps it
func (s *sessionService) SearchKnowledge(ctx context.Context,
	knowledgeBaseIDs []string, knowledgeIDs []string, tagScopes []types.TagScope, query string,
	opts *types.KnowledgeSearchOptions,
) (*types.RetrievalResult, error) {
	logger.Info(ctx, "Start knowledge base search without LLM summary")
	logger.Infof(ctx, "Knowledge base search parameters, knowledge base IDs: %v, knowledge IDs: %v, tag scopes: %d, query: %s",
		knowledgeBaseIDs, knowledgeIDs, len(tagScopes), query)
	if opts == nil {
		opts = &types.KnowledgeSearchOptions{}
	}
	if opts.DisableVectorMatch && opts.DisableKeywordsMatch {
		return nil, apperrors.NewBadRequestError("disable_vector_match and disable_keywords_match cannot both be true")
	}

	// Get tenant ID from context
	tenantID, ok := types.TenantIDFromContext(ctx)
	if !ok {
		logger.Error(ctx, "Failed to get tenant ID from context")
		return nil, fmt.Errorf("workspace ID not found in context")
	}

	// Load tenant-level retrieval config (nil is safe — GetEffective* methods handle nil receiver)
	var rc *types.RetrievalConfig
	if tenant, err2 := s.tenantService.GetTenantByID(ctx, tenantID); err2 == nil {
		rc = tenant.RetrievalConfig
	}

	// Resolve the rerank model before retrieving so a bad model_id costs
	// nothing. No rerank object means "rerank with the tenant config", which
	// is what this endpoint has always done.
	rerankDisabled := opts.Rerank != nil && !opts.Rerank.IsEnabled()
	var rerankModelID, rerankModelSource string
	if !rerankDisabled {
		requested := ""
		if opts.Rerank != nil {
			requested = opts.Rerank.ModelID
		}
		var err error
		rerankModelID, rerankModelSource, err = resolveRerankModelID(ctx, s.modelService, requested, rc)
		if err != nil {
			return nil, err
		}
	}
	// A model that cannot be loaded degrades to the retrieval order, as on
	// hybrid-search, instead of failing the search in the rerank stage.
	var unavailable *types.RerankDiagnostics
	if rerankModelID != "" {
		if _, err := s.modelService.GetRerankModel(ctx, rerankModelID); err != nil {
			logger.Warnf(ctx, "Rerank model %s unavailable, searching without rerank: %v", rerankModelID, err)
			unavailable = &types.RerankDiagnostics{
				Outcome: types.RerankOutcomeModelUnavailable,
				ModelID: rerankModelID,
				Error:   err.Error(),
			}
			rerankModelID = ""
		}
	}

	// Build unified search targets (computed once, used throughout pipeline)
	searchTargets, err := s.buildSearchTargets(ctx, tenantID, knowledgeBaseIDs, knowledgeIDs, tagScopes)
	if err != nil {
		return nil, fmt.Errorf("build search targets: %w", err)
	}

	chatManage := &types.ChatManage{
		PipelineRequest: types.PipelineRequest{
			Query:                query,
			UserID:               types.SessionOwnerIDFromContext(ctx),
			KnowledgeBaseIDs:     knowledgeBaseIDs,
			KnowledgeIDs:         knowledgeIDs,
			SearchTargets:        searchTargets,
			MaxRounds:            s.cfg.Conversation.MaxRounds,
			EmbeddingTopK:        rc.GetEffectiveEmbeddingTopK(),
			VectorThreshold:      rc.GetEffectiveVectorThreshold(),
			KeywordThreshold:     rc.GetEffectiveKeywordThreshold(),
			DisableVectorMatch:   opts.DisableVectorMatch,
			DisableKeywordsMatch: opts.DisableKeywordsMatch,
			RerankModelID:        rerankModelID,
			RerankTopK:           rc.GetEffectiveRerankTopK(),
			RerankThreshold:      rc.GetEffectiveRerankThreshold(),
		},
		PipelineState: types.PipelineState{
			RewriteQuery: query,
		},
	}
	applyKnowledgeSearchOverrides(chatManage, opts)

	results, err := s.runKnowledgeSearchPipeline(ctx, chatManage, len(searchTargets))
	if err != nil {
		return nil, err
	}

	diag := chatManage.RerankDiagnostics
	switch {
	case rerankDisabled:
		diag = &types.RerankDiagnostics{Outcome: types.RerankOutcomeDisabled}
	case unavailable != nil:
		diag = unavailable
		diag.Threshold, diag.EffectiveThreshold = chatManage.RerankThreshold, chatManage.RerankThreshold
	case diag == nil:
		// Retrieval found nothing, so the rerank stage never ran.
		diag = &types.RerankDiagnostics{
			Outcome:            types.RerankOutcomeNoCandidates,
			ModelID:            rerankModelID,
			Threshold:          chatManage.RerankThreshold,
			EffectiveThreshold: chatManage.RerankThreshold,
		}
	}
	diag.ModelSource = rerankModelSource
	if rerankDisabled {
		diag.ModelSource = ""
	}
	if rerankDisabled || unavailable != nil {
		// The rerank stage did not run; the results are the retrieval order.
		diag.CandidateCount = len(results)
	}
	diag.ResultCount = len(results)

	logger.Infof(ctx, "Knowledge base search completed, found %d results, rerank outcome: %s",
		len(results), diag.Outcome)
	return &types.RetrievalResult{Results: results, Meta: types.RetrievalMeta{Rerank: diag}}, nil
}

// applyKnowledgeSearchOverrides applies the caller's knowledge-search
// overrides on top of the tenant RetrievalConfig defaults in chatManage.
func applyKnowledgeSearchOverrides(chatManage *types.ChatManage, opts *types.KnowledgeSearchOptions) {
	if opts.VectorThreshold != nil {
		chatManage.VectorThreshold = *opts.VectorThreshold
	}
	if opts.KeywordThreshold != nil {
		chatManage.KeywordThreshold = *opts.KeywordThreshold
	}
	// The final FILTER_TOP_K stage cuts to RerankTopK whether or not the
	// rerank stage runs, so it is the result count. Recall must be at least
	// as deep for the count to be reachable.
	if opts.MatchCount > 0 {
		chatManage.RerankTopK = opts.MatchCount
	}
	if opts.Rerank != nil {
		if opts.Rerank.TopK > 0 {
			chatManage.RerankTopK = opts.Rerank.TopK
		}
		if opts.Rerank.Threshold != nil {
			chatManage.RerankThreshold = *opts.Rerank.Threshold
		}
	}
	chatManage.EmbeddingTopK = max(chatManage.EmbeddingTopK, chatManage.RerankTopK)
}

// runKnowledgeSearchPipeline runs the retrieval-only pipeline stages and
// returns the merged results. An empty search or an all-rejecting rerank is
// an empty result, not an error.
func (s *sessionService) runKnowledgeSearchPipeline(
	ctx context.Context,
	chatManage *types.ChatManage,
	targetCount int,
) ([]*types.SearchResult, error) {
	if targetCount == 0 {
		logger.Warn(ctx, "No search targets available, returning empty results")
		return []*types.SearchResult{}, nil
	}

	// Use specific event list, only including retrieval-related events, not LLM summarization
	searchEvents := []types.EventType{
		types.CHUNK_SEARCH, // Vector search
		types.CHUNK_RERANK, // Rerank search results
		types.CHUNK_MERGE,  // Merge search results
		types.FILTER_TOP_K, // Filter top K results
	}

	logger.Infof(ctx, "Trigger search event list: %v", searchEvents)

	for _, event := range searchEvents {
		logger.Infof(ctx, "Starting to trigger search event: %v", event)
		stageCtx, stageSpan := langfuse.GetManager().StartSpan(ctx, langfuse.SpanOptions{
			Name: "pipeline." + string(event),
			Metadata: map[string]interface{}{
				"event_type": string(event),
				"flow":       "search_knowledge",
			},
		})
		err := s.eventManager.Trigger(stageCtx, event, chatManage)
		var spanErr error
		if err != nil && err != chatpipeline.ErrSearchNothing {
			spanErr = err.Err
		}
		stageSpan.Finish(nil, nil, spanErr)

		if err == chatpipeline.ErrSearchNothing {
			logger.Warnf(ctx, "Event %v triggered, search result is empty", event)
			return []*types.SearchResult{}, nil
		}

		if err != nil {
			logger.Errorf(ctx, "Event triggering failed, event: %v, error type: %s, description: %s, error: %v",
				event, err.ErrorType, err.Description, err.Err)
			return nil, err.Err
		}
		logger.Infof(ctx, "Event %v triggered successfully", event)
	}
	return chatManage.MergeResult, nil
}

// handleFallbackResponse handles fallback response based on strategy
func (s *sessionService) handleFallbackResponse(ctx context.Context, chatManage *types.ChatManage) {
	if chatManage.FallbackStrategy == types.FallbackStrategyModel {
		s.handleModelFallback(ctx, chatManage)
	} else {
		s.handleFixedFallback(ctx, chatManage)
	}
}

// handleFixedFallback handles fixed fallback response
func (s *sessionService) handleFixedFallback(ctx context.Context, chatManage *types.ChatManage) {
	fallbackContent := chatManage.FallbackResponse
	chatManage.ChatResponse = &types.ChatResponse{Content: fallbackContent}
	s.emitFallbackAnswer(ctx, chatManage, fallbackContent)
}

// handleModelFallback handles model-based fallback response using streaming
func (s *sessionService) handleModelFallback(ctx context.Context, chatManage *types.ChatManage) {
	// Check if FallbackPrompt is available
	if chatManage.FallbackPrompt == "" {
		logger.Warnf(ctx, "Fallback strategy is 'model' but FallbackPrompt is empty, falling back to fixed response")
		s.handleFixedFallback(ctx, chatManage)
		return
	}

	// Render template with Query variable
	promptContent, err := s.renderFallbackPrompt(ctx, chatManage)
	if err != nil {
		logger.Errorf(ctx, "Failed to render fallback prompt: %v, falling back to fixed response", err)
		s.handleFixedFallback(ctx, chatManage)
		return
	}

	// Check if EventBus is available for streaming
	if chatManage.EventBus == nil {
		logger.Warnf(ctx, "EventBus not available for streaming fallback, falling back to fixed response")
		s.handleFixedFallback(ctx, chatManage)
		return
	}

	// Get chat model
	chatModel, err := s.modelService.GetChatModel(chatManage.ChatModelContext(ctx), chatManage.ChatModelID)
	if err != nil {
		logger.Errorf(ctx, "Failed to get chat model for fallback: %v, falling back to fixed response", err)
		s.handleFixedFallback(ctx, chatManage)
		return
	}

	// Prepare chat options
	thinking := false
	opt := &chat.ChatOptions{
		Temperature:         chatManage.SummaryConfig.Temperature,
		MaxCompletionTokens: chatManage.SummaryConfig.MaxCompletionTokens,
		Thinking:            &thinking,
	}

	// Start streaming response
	fallbackMessages, modelContext := prepareFallbackMessages(chatManage, promptContent)
	if leaks := modelContext.LeakedIdentifiers(fallbackMessages); len(leaks) > 0 {
		logger.Warnf(ctx, "[Fallback][ModelContext] %d message field(s) carry raw identifiers after encoding: %s",
			len(leaks), modelcontext.SummarizeLeaks(leaks))
	}
	responseChan, err := chatModel.ChatStream(ctx, fallbackMessages, opt)
	if err != nil {
		logger.Errorf(ctx, "Failed to start streaming fallback response: %v, falling back to fixed response", err)
		s.handleFixedFallback(ctx, chatManage)
		return
	}

	if responseChan == nil {
		logger.Errorf(ctx, "Chat stream returned nil channel, falling back to fixed response")
		s.handleFixedFallback(ctx, chatManage)
		return
	}

	// Start goroutine to consume stream and emit events
	go s.consumeFallbackStream(ctx, chatManage, responseChan, modelContext)
}

func prepareFallbackMessages(
	chatManage *types.ChatManage,
	promptContent string,
) ([]chat.Message, *modelcontext.Registry) {
	messages := buildFallbackMessages(chatManage, promptContent)
	citationsEnabled := chatManage == nil || chatManage.CitationsEnabled()
	registry := modelcontext.NewRegistry(citationsEnabled)
	if len(messages) > 0 && messages[0].Role == "system" {
		messages[0].Content = strings.TrimRight(messages[0].Content, " \t\r\n") + registry.ProtocolPrompt()
	} else {
		messages = append([]chat.Message{{Role: "system", Content: strings.TrimSpace(registry.ProtocolPrompt())}}, messages...)
	}
	return registry.EncodeMessages(messages), registry
}

func buildFallbackMessages(chatManage *types.ChatManage, promptContent string) []chat.Message {
	messages := make([]chat.Message, 0, len(chatManage.History)*2+2)

	// The model-fallback prompt is a system-style instruction (KB document
	// listing + "use general knowledge when nothing matched" guidance). Carry
	// it in the system role so the LLM input keeps a proper system message
	// instead of starting with a bare user turn. We deliberately do NOT reuse
	// the RAG summary system prompt (SummaryConfig.Prompt) here: that template
	// forbids prior knowledge ("reply ONLY based on retrieved information"),
	// which directly contradicts the fallback's purpose.
	if strings.TrimSpace(promptContent) != "" {
		messages = append(messages, chat.Message{
			Role: "system",
			Content: promptContent + "\n\n" + types.SourceDataBoundaryPrompt +
				"\n\n" + types.SourcedAnswerOutputPrompt,
		})
	}

	messages = chatpipeline.AppendHistoryMessages(messages, chatManage.History)

	// End on the user's actual question so generation is prompted by a user
	// turn (the query is also embedded in the system instruction above, but a
	// trailing user message keeps the chat shape valid for all providers).
	query := chatManage.Query
	if rq := strings.TrimSpace(chatManage.RewriteQuery); rq != "" {
		query = rq
	}
	userMsg := chat.Message{Role: "user", Content: query}
	if chatManage.ChatModelSupportsVision && len(chatManage.Images) > 0 {
		userMsg.Images = chatManage.Images
	}

	return append(messages, userMsg)
}

// renderFallbackPrompt renders the fallback prompt template with query and image context.
func (s *sessionService) renderFallbackPrompt(ctx context.Context, chatManage *types.ChatManage) (string, error) {
	query := chatManage.Query
	if rq := strings.TrimSpace(chatManage.RewriteQuery); rq != "" {
		query = rq
	}

	kbDocuments := s.buildKBDocumentListing(ctx, chatManage)

	result := types.RenderPromptPlaceholders(chatManage.FallbackPrompt, types.PlaceholderValues{
		"query":        query,
		"language":     chatManage.Language,
		"kb_documents": kbDocuments,
	})

	if chatManage.ImageDescription != "" && !chatManage.ChatModelSupportsVision {
		result += "\n\n[用户上传图片内容]\n" + chatManage.ImageDescription
	}
	if chatManage.QuotedContext != "" {
		result += "\n\n" + chatManage.QuotedContext
	}
	return result, nil
}

// buildKBDocumentListing returns a concise listing of documents in the knowledge bases
// associated with the current pipeline. This gives the LLM visibility into KB contents
// when vector/keyword search returns empty (e.g., broad browse queries).
func (s *sessionService) buildKBDocumentListing(ctx context.Context, chatManage *types.ChatManage) string {
	// Collect unique KB IDs from search targets
	kbIDs := make(map[string]struct{})
	for _, t := range chatManage.SearchTargets {
		kbIDs[t.KnowledgeBaseID] = struct{}{}
	}
	for _, id := range chatManage.KnowledgeBaseIDs {
		kbIDs[id] = struct{}{}
	}
	if len(kbIDs) == 0 {
		return ""
	}

	const maxDocuments = 50
	var b strings.Builder
	total := 0

	for kbID := range kbIDs {
		if total >= maxDocuments {
			break
		}
		knowledges, err := s.knowledgeService.ListKnowledgeByKnowledgeBaseID(ctx, kbID)
		if err != nil {
			logger.Warnf(ctx, "buildKBDocumentListing: failed to list knowledge for KB %s: %v", kbID, err)
			continue
		}
		for _, k := range knowledges {
			if total >= maxDocuments {
				break
			}
			if k.EnableStatus != "enabled" {
				continue
			}
			title := k.Title
			if title == "" {
				title = k.FileName
			}
			if title == "" {
				continue
			}
			fmt.Fprintf(&b, "- %s", title)
			if k.FileType != "" {
				fmt.Fprintf(&b, " (%s)", k.FileType)
			}
			if k.Description != "" {
				desc := k.Description
				if len([]rune(desc)) > 100 {
					desc = string([]rune(desc)[:100]) + "..."
				}
				fmt.Fprintf(&b, ": %s", desc)
			}
			b.WriteString("\n")
			total++
		}
	}

	if b.Len() == 0 {
		return ""
	}

	if total >= maxDocuments {
		fmt.Fprintf(&b, "... (showing first %d documents)\n", maxDocuments)
	}

	return b.String()
}

// consumeFallbackStream consumes the streaming response and emits events
func (s *sessionService) consumeFallbackStream(
	ctx context.Context,
	chatManage *types.ChatManage,
	responseChan <-chan types.StreamResponse,
	modelContext *modelcontext.Registry,
) {
	fallbackID := generateEventID("fallback")
	eventBus := chatManage.EventBus
	var finalContent string
	streamCompleted := false
	decoder := modelContext.StreamDecoder()

	for response := range responseChan {
		// Emit event for each answer chunk
		if response.ResponseType == types.ResponseTypeAnswer {
			response.Content = decoder.Feed(response.Content)
			if response.Done {
				response.Content += decoder.Flush()
			}
			truncated := response.Done && chatpipeline.IsLengthFinishReason(response.FinishReason)
			if truncated && strings.TrimSpace(finalContent+response.Content) == "" {
				response.Content = chatpipeline.EmptyTruncatedAnswerFallback
			}
			finalContent += response.Content
			if err := eventBus.Emit(ctx, types.Event{
				ID:        fallbackID,
				Type:      types.EventType(event.EventAgentFinalAnswer),
				SessionID: chatManage.SessionID,
				Data: event.AgentFinalAnswerData{
					Content:    response.Content,
					Done:       response.Done,
					IsFallback: true,
					Truncated:  truncated,
				},
			}); err != nil {
				logger.Errorf(ctx, "Failed to emit fallback answer chunk event: %v", err)
			}

			// Update ChatResponse with final content when done
			if response.Done {
				chatManage.ChatResponse = &types.ChatResponse{Content: finalContent}
				streamCompleted = true
				logger.Infof(ctx, "Fallback streaming response completed")
				break
			}
		}
	}

	// If channel closed without Done=true, emit final event with fixed response
	if !streamCompleted {
		logger.Warnf(ctx, "Fallback stream closed without completion, emitting final event with fixed response")
		s.emitFallbackAnswer(ctx, chatManage, chatManage.FallbackResponse)
	}
}

// emitKnowledgeReferencesEvent streams retrieved chunks to the client as a
// `references` SSE event. These references drive the retrieval-results UI and
// remain available even when inline citations in the model answer are disabled.
// Must run before CHAT_COMPLETION_STREAM so the event arrives while the
// connection is still open (complete closes the stream).
func emitKnowledgeReferencesEvent(ctx context.Context, chatManage *types.ChatManage) {
	if chatManage == nil || chatManage.EventBus == nil || len(chatManage.MergeResult) == 0 {
		return
	}
	logger.Infof(ctx, "Emitting references event with %d results (pre-answer)", len(chatManage.MergeResult))
	if err := chatManage.EventBus.Emit(ctx, types.Event{
		ID:        generateEventID("references"),
		Type:      types.EventType(event.EventAgentReferences),
		SessionID: chatManage.SessionID,
		Data: event.AgentReferencesData{
			References: chatManage.MergeResult,
		},
	}); err != nil {
		logger.Errorf(ctx, "Failed to emit references event: %v", err)
	}
}

// emitFallbackAnswer emits fallback answer event
func (s *sessionService) emitFallbackAnswer(ctx context.Context, chatManage *types.ChatManage, content string) {
	if chatManage.EventBus == nil {
		return
	}
	if !chatManage.CitationsEnabled() {
		content = modelcontext.NewRegistry(false).DecodeOutputText(content)
	}

	fallbackID := generateEventID("fallback")
	if err := chatManage.EventBus.Emit(ctx, types.Event{
		ID:        fallbackID,
		Type:      types.EventType(event.EventAgentFinalAnswer),
		SessionID: chatManage.SessionID,
		Data: event.AgentFinalAnswerData{
			Content:    content,
			Done:       true,
			IsFallback: true,
		},
	}); err != nil {
		logger.Errorf(ctx, "Failed to emit fallback answer event: %v", err)
	} else {
		logger.Infof(ctx, "Fallback answer event emitted successfully")
	}
}

// resolveWebSearchProviderID returns the web search provider ID to use for a pipeline request.
// Priority: agent config > tenant default (is_default=true)
func (s *sessionService) resolveWebSearchProviderID(ctx context.Context, req *types.QARequest, tenantID uint64) string {
	// 1. Agent-level override
	if req.CustomAgent != nil && req.CustomAgent.Config.WebSearchProviderID != "" {
		return req.CustomAgent.Config.WebSearchProviderID
	}
	// 2. Tenant default
	if s.webSearchProviderRepo != nil {
		if defaultProvider, err := s.webSearchProviderRepo.GetDefault(ctx, tenantID); err == nil && defaultProvider != nil {
			return defaultProvider.ID
		}
	}
	return ""
}

// resolveWebSearchEnabled combines the request switch with the agent's own
// setting, as agent mode does: the switch only opts a turn in, so a client
// cannot turn on a search (and spend the agent workspace's provider quota)
// that the agent disables. Requests without an agent keep the switch.
func resolveWebSearchEnabled(req *types.QARequest) bool {
	if req.CustomAgent != nil {
		return req.CustomAgent.Config.WebSearchEnabled && req.WebSearchEnabled
	}
	return req.WebSearchEnabled
}

// resolveWebFetchEnabled returns whether auto web fetch is enabled for this request.
func (s *sessionService) resolveWebFetchEnabled(req *types.QARequest) bool {
	if req.CustomAgent != nil {
		return req.CustomAgent.Config.WebFetchEnabled
	}
	return false
}

// resolveWebFetchTopN returns how many pages to fetch after rerank.
func (s *sessionService) resolveWebFetchTopN(req *types.QARequest) int {
	if req.CustomAgent != nil && req.CustomAgent.Config.WebFetchTopN > 0 {
		return req.CustomAgent.Config.WebFetchTopN
	}
	return 3
}

// resolveWebSearchMaxResults returns the max results for web search.
// Priority: agent config > tenant default > default (10)
func (s *sessionService) resolveWebSearchMaxResults(ctx context.Context, req *types.QARequest) int {
	if req.CustomAgent != nil && req.CustomAgent.Config.WebSearchMaxResults > 0 {
		return req.CustomAgent.Config.WebSearchMaxResults
	}
	tenantInfo, _ := types.TenantInfoFromContext(ctx)
	if tenantInfo != nil {
		return types.EffectiveWebSearchConfig(tenantInfo.WebSearchConfig).MaxResults
	}
	return types.DefaultWebSearchMaxResults
}