diff --git a/Makefile b/Makefile
index 9503bf7b4..cafcdd44a 100644
--- a/Makefile
+++ b/Makefile
@@ -180,7 +180,7 @@ osx-signed: build
 
 ## Run
 run: ## run local-ai
-	CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GOCMD) run ./
+	CGO_LDFLAGS="$(CGO_LDFLAGS)" $(GOCMD) run ./cmd/local-ai
 
 prepare-test: protogen-go build-mock-backend
 
diff --git a/core/services/agentpool/agent_pool.go b/core/services/agentpool/agent_pool.go
index 9216add9b..8744b8a35 100644
--- a/core/services/agentpool/agent_pool.go
+++ b/core/services/agentpool/agent_pool.go
@@ -466,10 +466,11 @@ func (s *AgentPoolService) Chat(name, message string) (string, error) {
 				s.collectAndCopyMetadata(metadata, chatUserID)
 			}
 
+			content := s.appendLocalAGIKBCitations(response.Response, name, message, response.State)
 			msg := map[string]any{
 				"id":        messageID + "-agent",
 				"sender":    "agent",
-				"content":   response.Response,
+				"content":   content,
 				"timestamp": time.Now().Format(time.RFC3339),
 			}
 			if len(metadata) > 0 {
@@ -489,6 +490,79 @@ func (s *AgentPoolService) Chat(name, message string) (string, error) {
 	return messageID, nil
 }
 
+func (s *AgentPoolService) appendLocalAGIKBCitations(response, agentKey, message string, states []coreTypes.ActionState) string {
+	if strings.TrimSpace(response) == "" {
+		return response
+	}
+
+	userID, collection := splitAgentKey(agentKey)
+	cfg := s.localAGI.pool.GetConfig(agentKey)
+	if cfg == nil || !cfg.EnableKnowledgeBase {
+		return response
+	}
+
+	citations := kbCitationsFromActionStates(states)
+	if len(citations) == 0 && cfg.KBAutoSearch {
+		maxResults := cfg.KnowledgeBaseResults
+		if maxResults <= 0 {
+			maxResults = 5
+		}
+		ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+		defer cancel()
+		kbResult := agents.KBAutoSearchPrompt(ctx, s.apiURL, s.apiKey, collection, message, maxResults, userID)
+		citations = kbResult.Citations
+	}
+
+	return agents.AppendKBCitations(response, collection, userID, citations)
+}
+
+func splitAgentKey(agentKey string) (userID, name string) {
+	if uid, n, ok := strings.Cut(agentKey, ":"); ok {
+		return uid, n
+	}
+	return "", agentKey
+}
+
+func kbCitationsFromActionStates(states []coreTypes.ActionState) []agents.KBCitation {
+	var citations []agents.KBCitation
+	for _, state := range states {
+		citations = append(citations, kbCitationsFromMetadata(state.Metadata)...)
+	}
+	return citations
+}
+
+func kbCitationsFromMetadata(metadata map[string]any) []agents.KBCitation {
+	if len(metadata) == 0 {
+		return nil
+	}
+
+	fileName := metadata["file_name"]
+	source := metadata["source"]
+	if fileName == nil && source == nil {
+		return nil
+	}
+
+	citation := agents.KBCitation{
+		FileName: metadataString(fileName),
+		EntryKey: metadataString(source),
+	}
+	if citation.FileName == "" && citation.EntryKey == "" {
+		return nil
+	}
+	return []agents.KBCitation{citation}
+}
+
+func metadataString(value any) string {
+	switch v := value.(type) {
+	case string:
+		return v
+	case fmt.Stringer:
+		return v.String()
+	default:
+		return ""
+	}
+}
+
 // userOutputsDir returns the per-user outputs directory, creating it if needed.
 // If userID is empty, falls back to the shared outputs directory.
 func (s *AgentPoolService) userOutputsDir(userID string) string {
diff --git a/core/services/agents/citations.go b/core/services/agents/citations.go
new file mode 100644
index 000000000..65c96d75c
--- /dev/null
+++ b/core/services/agents/citations.go
@@ -0,0 +1,127 @@
+package agents
+
+import (
+	"fmt"
+	"net/url"
+	"strings"
+	"sync"
+)
+
+type kbCitationList struct {
+	mu        sync.Mutex
+	citations []KBCitation
+}
+
+func (l *kbCitationList) AddKBCitations(citations []KBCitation) {
+	if len(citations) == 0 {
+		return
+	}
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	l.citations = append(l.citations, citations...)
+}
+
+func (l *kbCitationList) Citations() []KBCitation {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	out := make([]KBCitation, len(l.citations))
+	copy(out, l.citations)
+	return out
+}
+
+// AppendKBCitations appends a markdown Sources block for KB citations.
+func AppendKBCitations(response, collection, userID string, citations []KBCitation) string {
+	if strings.TrimSpace(response) == "" || len(citations) == 0 {
+		return response
+	}
+
+	var lines []string
+	seen := make(map[string]struct{})
+	for _, citation := range citations {
+		key := strings.TrimSpace(citation.EntryKey)
+		if key == "" {
+			key = strings.TrimSpace(citation.FileName)
+		}
+		if key == "" {
+			continue
+		}
+		if _, ok := seen[key]; ok {
+			continue
+		}
+		seen[key] = struct{}{}
+
+		displayName := kbCitationDisplayName(citation)
+		if displayName == "" {
+			continue
+		}
+
+		sourceURL := kbCitationRawFileURL(collection, citation.EntryKey, userID)
+		number := len(lines) + 1
+		if sourceURL == "" {
+			lines = append(lines, fmt.Sprintf("[%d] %s", number, displayName))
+			continue
+		}
+		lines = append(lines, fmt.Sprintf("[%d] [%s](%s)", number, escapeMarkdownLinkText(displayName), sourceURL))
+	}
+
+	if len(lines) == 0 {
+		return response
+	}
+
+	var sb strings.Builder
+	sb.WriteString(strings.TrimRight(response, "\n"))
+	sb.WriteString("\n\nSources:\n")
+	for _, line := range lines {
+		sb.WriteString(line)
+		sb.WriteString("\n")
+	}
+	return strings.TrimRight(sb.String(), "\n")
+}
+
+func kbCitationDisplayName(citation KBCitation) string {
+	if fileName := strings.TrimSpace(citation.FileName); fileName != "" {
+		return fileName
+	}
+
+	segments := strings.Split(strings.Trim(strings.TrimSpace(citation.EntryKey), "/"), "/")
+	for i := len(segments) - 1; i >= 0; i-- {
+		if segment := strings.TrimSpace(segments[i]); segment != "" {
+			return segment
+		}
+	}
+	return ""
+}
+
+func kbCitationRawFileURL(collection, entryKey, userID string) string {
+	collection = strings.TrimSpace(collection)
+	entryKey = strings.Trim(strings.TrimSpace(entryKey), "/")
+	if collection == "" || entryKey == "" {
+		return ""
+	}
+
+	var escapedEntrySegments []string
+	for _, segment := range strings.Split(entryKey, "/") {
+		if segment == "" {
+			continue
+		}
+		escapedEntrySegments = append(escapedEntrySegments, url.PathEscape(segment))
+	}
+	if len(escapedEntrySegments) == 0 {
+		return ""
+	}
+
+	sourceURL := "/api/agents/collections/" + url.PathEscape(collection) + "/entries-raw/" + strings.Join(escapedEntrySegments, "/")
+	if userID != "" {
+		query := url.Values{}
+		query.Set("user_id", userID)
+		sourceURL += "?" + query.Encode()
+	}
+	return sourceURL
+}
+
+func escapeMarkdownLinkText(text string) string {
+	text = strings.ReplaceAll(text, `\`, `\\`)
+	text = strings.ReplaceAll(text, "[", `\[`)
+	text = strings.ReplaceAll(text, "]", `\]`)
+	return text
+}
diff --git a/core/services/agents/executor.go b/core/services/agents/executor.go
index 81481600c..2787aeabc 100644
--- a/core/services/agents/executor.go
+++ b/core/services/agents/executor.go
@@ -167,10 +167,12 @@ func ExecuteChatWithLLM(ctx context.Context, llm cogito.LLM, cfg *AgentConfig, m
 		}
 	}
 
+	kbCitations := &kbCitationList{}
 	if cfg.EnableKnowledgeBase && (kbMode == KBModeAutoSearch || kbMode == KBModeBoth) {
-		kbResults := KBAutoSearchPrompt(ctx, effectiveURL, effectiveKey, cfg.Name, message, cfg.KnowledgeBaseResults, userID)
-		if kbResults != "" {
-			fragment = fragment.AddMessage(cogito.SystemMessageRole, kbResults)
+		kbResult := KBAutoSearchPrompt(ctx, effectiveURL, effectiveKey, cfg.Name, message, cfg.KnowledgeBaseResults, userID)
+		if kbResult.Prompt != "" {
+			fragment = fragment.AddMessage(cogito.SystemMessageRole, kbResult.Prompt)
+			kbCitations.AddKBCitations(kbResult.Citations)
 		}
 	}
 
@@ -197,7 +199,7 @@ func ExecuteChatWithLLM(ctx context.Context, llm cogito.LLM, cfg *AgentConfig, m
 		}
 		cogitoOpts = append(cogitoOpts, cogito.WithTools(
 			cogito.NewToolDefinition(
-				KBSearchMemoryTool{APIURL: effectiveURL, APIKey: effectiveKey, Collection: cfg.Name, MaxResults: kbResults, UserID: userID},
+				KBSearchMemoryTool{APIURL: effectiveURL, APIKey: effectiveKey, Collection: cfg.Name, MaxResults: kbResults, UserID: userID, CitationCollector: kbCitations},
 				KBSearchMemoryArgs{},
 				"search_memory",
 				"Search the knowledge base for relevant information",
@@ -336,6 +338,8 @@ func ExecuteChatWithLLM(ctx context.Context, llm cogito.LLM, cfg *AgentConfig, m
 	if cfg.StripThinkingTags && response != "" {
 		response = stripThinkingTags(response)
 	}
+	responseForMemory := response
+	response = AppendKBCitations(response, cfg.Name, userID, kbCitations.Citations())
 
 	// Save conversation to KB when long-term memory is enabled.
 	// Use a detached context: the parent ctx may be cancelled (e.g. in distributed
@@ -344,7 +348,7 @@ func ExecuteChatWithLLM(ctx context.Context, llm cogito.LLM, cfg *AgentConfig, m
 		go func() {
 			ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
 			defer cancel()
-			saveConversationToKB(ctx, llm, effectiveURL, effectiveKey, cfg, message, response, userID)
+			saveConversationToKB(ctx, llm, effectiveURL, effectiveKey, cfg, message, responseForMemory, userID)
 		}()
 	}
 
diff --git a/core/services/agents/executor_test.go b/core/services/agents/executor_test.go
index 04d44462c..9b2001cca 100644
--- a/core/services/agents/executor_test.go
+++ b/core/services/agents/executor_test.go
@@ -2,6 +2,8 @@ package agents
 
 import (
 	"context"
+	"net/http"
+	"net/http/httptest"
 	"sync"
 	"sync/atomic"
 
@@ -36,6 +38,34 @@ func (m *mockLLM) CreateChatCompletion(ctx context.Context, req openai.ChatCompl
 	}, cogito.LLMUsage{}, nil
 }
 
+type toolCallingMockLLM struct {
+	createResponses []openai.ChatCompletionResponse
+	askResponse     string
+	callCount       atomic.Int32
+}
+
+func (m *toolCallingMockLLM) Ask(ctx context.Context, f cogito.Fragment) (cogito.Fragment, error) {
+	m.callCount.Add(1)
+	return f.AddMessage(cogito.AssistantMessageRole, m.askResponse), nil
+}
+
+func (m *toolCallingMockLLM) CreateChatCompletion(ctx context.Context, req openai.ChatCompletionRequest) (cogito.LLMReply, cogito.LLMUsage, error) {
+	idx := int(m.callCount.Add(1)) - 1
+	if idx >= len(m.createResponses) {
+		return cogito.LLMReply{
+			ChatCompletionResponse: openai.ChatCompletionResponse{
+				Choices: []openai.ChatCompletionChoice{{
+					Message: openai.ChatCompletionMessage{
+						Role:    "assistant",
+						Content: "No more tools needed.",
+					},
+				}},
+			},
+		}, cogito.LLMUsage{}, nil
+	}
+	return cogito.LLMReply{ChatCompletionResponse: m.createResponses[idx]}, cogito.LLMUsage{}, nil
+}
+
 // statusCollector records status callbacks in a thread-safe way.
 type statusCollector struct {
 	mu       sync.Mutex
@@ -73,6 +103,74 @@ var _ = DescribeTable("stripThinkingTags",
 	Entry("adjacent tag pairs", "<thinking>a</thinking><thinking>b</thinking>", ""),
 )
 
+var _ = DescribeTable("appendKBCitations",
+	func(response, collection, userID string, citations []KBCitation, want string) {
+		Expect(AppendKBCitations(response, collection, userID, citations)).To(Equal(want))
+	},
+	Entry("leaves responses without citations unchanged",
+		"answer",
+		"agent",
+		"",
+		nil,
+		"answer",
+	),
+	Entry("leaves blank responses unchanged",
+		"",
+		"agent",
+		"",
+		[]KBCitation{{FileName: "source.pdf", EntryKey: "uuid/source.pdf"}},
+		"",
+	),
+	Entry("appends clickable source links",
+		"answer",
+		"my-agent",
+		"",
+		[]KBCitation{{FileName: "new feature.pdf", EntryKey: "uuid/new feature.pdf"}},
+		"answer\n\nSources:\n[1] [new feature.pdf](/api/agents/collections/my-agent/entries-raw/uuid/new%20feature.pdf)",
+	),
+	Entry("deduplicates citations by entry key",
+		"answer",
+		"agent",
+		"",
+		[]KBCitation{
+			{FileName: "first.pdf", EntryKey: "uuid/shared.pdf"},
+			{FileName: "second.pdf", EntryKey: "uuid/shared.pdf"},
+		},
+		"answer\n\nSources:\n[1] [first.pdf](/api/agents/collections/agent/entries-raw/uuid/shared.pdf)",
+	),
+	Entry("uses plain text when entry key is missing",
+		"answer",
+		"agent",
+		"",
+		[]KBCitation{{FileName: "source.pdf"}},
+		"answer\n\nSources:\n[1] source.pdf",
+	),
+	Entry("uses entry basename when filename is missing",
+		"answer",
+		"agent",
+		"",
+		[]KBCitation{{EntryKey: "uuid/source.pdf"}},
+		"answer\n\nSources:\n[1] [source.pdf](/api/agents/collections/agent/entries-raw/uuid/source.pdf)",
+	),
+	Entry("adds user id query when present",
+		"answer",
+		"agent",
+		"user 1",
+		[]KBCitation{{FileName: "source.pdf", EntryKey: "uuid/source.pdf"}},
+		"answer\n\nSources:\n[1] [source.pdf](/api/agents/collections/agent/entries-raw/uuid/source.pdf?user_id=user+1)",
+	),
+	Entry("escapes collection, path segments, and markdown link text",
+		"answer",
+		"agent one",
+		"",
+		[]KBCitation{{FileName: "source [draft].pdf", EntryKey: "uuid/source [draft].pdf"}},
+		`answer
+
+Sources:
+[1] [source \[draft\].pdf](/api/agents/collections/agent%20one/entries-raw/uuid/source%20%5Bdraft%5D.pdf)`,
+	),
+)
+
 var _ = Describe("ExecuteChatWithLLM", func() {
 	var (
 		ctx context.Context
@@ -184,6 +282,150 @@ var _ = Describe("ExecuteChatWithLLM", func() {
 		})
 	})
 
+	Context("knowledge base citations", func() {
+		It("appends KB sources to the returned response and callback message", func() {
+			mux := http.NewServeMux()
+			mux.HandleFunc("/api/agents/collections/kb-agent/search", func(w http.ResponseWriter, r *http.Request) {
+				Expect(r.URL.Query().Get("user_id")).To(Equal("user-1"))
+				w.Header().Set("Content-Type", "application/json")
+				_, _ = w.Write([]byte(`{
+					"results": [
+						{
+							"content": "KB content",
+							"id": "result-1",
+							"similarity": 0.99,
+							"metadata": {
+								"file_name": "new feature.pdf",
+								"source": "uuid/new feature.pdf"
+							}
+						}
+					],
+					"count": 1
+				}`))
+			})
+			server := httptest.NewServer(mux)
+			defer server.Close()
+
+			var msgContent string
+			cb.OnMessage = func(sender, content, messageID string) {
+				msgContent = content
+			}
+
+			llm := &mockLLM{response: "agent reply"}
+			cfg := &AgentConfig{
+				Name:                "kb-agent",
+				Model:               "test-model",
+				EnableKnowledgeBase: true,
+				KBMode:              KBModeAutoSearch,
+			}
+
+			result, err := ExecuteChatWithLLM(ctx, llm, cfg, "hello", cb, ExecuteChatOpts{
+				APIURL: server.URL,
+				UserID: "user-1",
+			})
+			Expect(err).ToNot(HaveOccurred())
+			Expect(result).To(Equal("agent reply\n\nSources:\n[1] [new feature.pdf](/api/agents/collections/kb-agent/entries-raw/uuid/new%20feature.pdf?user_id=user-1)"))
+			Expect(msgContent).To(Equal(result))
+		})
+
+		It("collects citations from the search_memory tool", func() {
+			mux := http.NewServeMux()
+			mux.HandleFunc("/api/agents/collections/kb-agent/search", func(w http.ResponseWriter, r *http.Request) {
+				w.Header().Set("Content-Type", "application/json")
+				_, _ = w.Write([]byte(`{
+					"results": [
+						{
+							"content": "Tool KB content",
+							"id": "result-1",
+							"similarity": 0.99,
+							"metadata": {
+								"file_name": "tool source.pdf",
+								"source": "uuid/tool source.pdf"
+							}
+						}
+					],
+					"count": 1
+				}`))
+			})
+			server := httptest.NewServer(mux)
+			defer server.Close()
+
+			collector := &kbCitationList{}
+			tool := KBSearchMemoryTool{
+				APIURL:            server.URL,
+				Collection:        "kb-agent",
+				CitationCollector: collector,
+			}
+
+			result, _, err := tool.Run(KBSearchMemoryArgs{Query: "hello"})
+			Expect(err).ToNot(HaveOccurred())
+			Expect(result).To(ContainSubstring("Tool KB content"))
+			Expect(collector.Citations()).To(Equal([]KBCitation{{FileName: "tool source.pdf", EntryKey: "uuid/tool source.pdf"}}))
+		})
+
+		It("appends KB sources found through tools-only search_memory calls", func() {
+			mux := http.NewServeMux()
+			mux.HandleFunc("/api/agents/collections/kb-agent/search", func(w http.ResponseWriter, r *http.Request) {
+				Expect(r.URL.Query().Get("user_id")).To(Equal("user-1"))
+				w.Header().Set("Content-Type", "application/json")
+				_, _ = w.Write([]byte(`{
+					"results": [
+						{
+							"content": "Tool KB content",
+							"id": "result-1",
+							"similarity": 0.99,
+							"metadata": {
+								"file_name": "tool source.pdf",
+								"source": "uuid/tool source.pdf"
+							}
+						}
+					],
+					"count": 1
+				}`))
+			})
+			server := httptest.NewServer(mux)
+			defer server.Close()
+
+			llm := &toolCallingMockLLM{
+				askResponse: "agent reply from tool context",
+				createResponses: []openai.ChatCompletionResponse{
+					{
+						Choices: []openai.ChatCompletionChoice{
+							{
+								Message: openai.ChatCompletionMessage{
+									Role: "assistant",
+									ToolCalls: []openai.ToolCall{
+										{
+											ID:   "call-1",
+											Type: openai.ToolTypeFunction,
+											Function: openai.FunctionCall{
+												Name:      "search_memory",
+												Arguments: `{"query":"hello"}`,
+											},
+										},
+									},
+								},
+							},
+						},
+					},
+				},
+			}
+			cfg := &AgentConfig{
+				Name:                "kb-agent",
+				Model:               "test-model",
+				EnableKnowledgeBase: true,
+				KBMode:              KBModeTools,
+			}
+
+			result, err := ExecuteChatWithLLM(ctx, llm, cfg, "hello", cb, ExecuteChatOpts{
+				APIURL: server.URL,
+				UserID: "user-1",
+			})
+			Expect(err).ToNot(HaveOccurred())
+			Expect(result).To(Equal("agent reply from tool context\n\nSources:\n[1] [tool source.pdf](/api/agents/collections/kb-agent/entries-raw/uuid/tool%20source.pdf?user_id=user-1)"))
+		})
+	})
+
 	Context("context cancellation", func() {
 		It("returns an error when context is already cancelled", func() {
 			cancelledCtx, cancel := context.WithCancel(ctx)
diff --git a/core/services/agents/knowledge.go b/core/services/agents/knowledge.go
index 8c0afb012..ae87f560d 100644
--- a/core/services/agents/knowledge.go
+++ b/core/services/agents/knowledge.go
@@ -8,6 +8,7 @@ import (
 	"io"
 	"mime/multipart"
 	"net/http"
+	"net/url"
 	"strings"
 	"time"
 
@@ -17,10 +18,19 @@ import (
 	"github.com/mudler/LocalAI/pkg/httpclient"
 )
 
+// Metadata keys populated by localrecall for every stored chunk. The original
+// upload file name lives under file_name (used for display); source holds the
+// collection entry key ("<uuid>/<filename>") used to build the raw-file URL.
+const (
+	kbMetadataFileName = "file_name"
+	kbMetadataSource   = "source"
+)
+
 // KBSearchResult represents a search result from the knowledge base.
+// Field names mirror the collection search endpoint's JSON response.
 type KBSearchResult struct {
 	Content    string            `json:"content"`
-	Score      float64           `json:"score"`
+	ID         string            `json:"id"`
 	Similarity float64           `json:"similarity"`
 	Metadata   map[string]string `json:"metadata"`
 }
@@ -31,22 +41,48 @@ type kbSearchResponse struct {
 	Count   int              `json:"count"`
 }
 
-// KBAutoSearchPrompt queries the knowledge base with the user's message
-// and returns a system prompt block with relevant results.
+// KBCitation is a single source document that a KB search drew from. Citations
+// travel alongside the prompt as structured data so the consumer (and UI) can
+// render clickable source links, independent of what the model writes inline.
+type KBCitation struct {
+	// FileName is the original uploaded file name, for display (e.g. "report.pdf").
+	FileName string `json:"file_name"`
+	// EntryKey is the collection entry identifier ("<uuid>/<filename>"), used to
+	// build the raw-file URL and as the de-duplication key.
+	EntryKey string `json:"entry_key"`
+}
+
+// KBSearchContext is the result of an auto-search against the knowledge base:
+// the system-prompt block to feed the model, plus the de-duplicated list of
+// source documents the results were drawn from.
+type KBSearchContext struct {
+	Prompt    string       `json:"prompt"`
+	Citations []KBCitation `json:"citations"`
+}
+
+// KBCitationCollector receives source citations found during KB searches.
+type KBCitationCollector interface {
+	AddKBCitations([]KBCitation)
+}
+
+// KBAutoSearchPrompt queries the knowledge base with the user's message and
+// returns a KBSearchContext: a system prompt block with the relevant results
+// plus the de-duplicated source citations those results came from.
 // Uses LocalAI's collection search endpoint via the API.
-func KBAutoSearchPrompt(ctx context.Context, apiURL, apiKey, collection, query string, maxResults int, userID string) string {
+func KBAutoSearchPrompt(ctx context.Context, apiURL, apiKey, collection, query string, maxResults int, userID string) KBSearchContext {
 	if collection == "" || query == "" {
-		return ""
+		return KBSearchContext{}
 	}
 
 	if maxResults <= 0 {
 		maxResults = 5
 	}
 
-	// Call LocalAI's collection search API
-	searchURL := strings.TrimRight(apiURL, "/") + "/api/agents/collections/" + collection + "/search"
+	searchURL := strings.TrimRight(apiURL, "/") + "/api/agents/collections/" + url.PathEscape(collection) + "/search"
 	if userID != "" {
-		searchURL += "?user_id=" + userID
+		query := url.Values{}
+		query.Set("user_id", userID)
+		searchURL += "?" + query.Encode()
 	}
 	reqBody, _ := json.Marshal(map[string]any{
 		"query":       query,
@@ -56,7 +92,7 @@ func KBAutoSearchPrompt(ctx context.Context, apiURL, apiKey, collection, query s
 	req, err := http.NewRequestWithContext(ctx, http.MethodPost, searchURL, strings.NewReader(string(reqBody)))
 	if err != nil {
 		xlog.Warn("KB auto-search: failed to create request", "error", err)
-		return ""
+		return KBSearchContext{}
 	}
 	req.Header.Set("Content-Type", "application/json")
 	if apiKey != "" {
@@ -66,41 +102,70 @@ func KBAutoSearchPrompt(ctx context.Context, apiURL, apiKey, collection, query s
 	resp, err := httpclient.New().Do(req)
 	if err != nil {
 		xlog.Warn("KB auto-search: request failed", "error", err)
-		return ""
+		return KBSearchContext{}
 	}
 	defer resp.Body.Close()
 
 	if resp.StatusCode != http.StatusOK {
 		body, _ := io.ReadAll(resp.Body)
 		xlog.Warn("KB auto-search: non-200 response", "status", resp.StatusCode, "body", string(body))
-		return ""
+		return KBSearchContext{}
 	}
 
 	var searchResp kbSearchResponse
 	if err := json.NewDecoder(resp.Body).Decode(&searchResp); err != nil {
 		xlog.Warn("KB auto-search: failed to decode response", "error", err)
-		return ""
+		return KBSearchContext{}
 	}
 
 	if len(searchResp.Results) == 0 {
-		return ""
+		return KBSearchContext{}
 	}
 
-	// Format results as a system prompt block (same format as LocalAGI)
+	// Build the system prompt block, labelling each chunk with its source file
+	// so the model can attribute inline, and collect the structured citations.
 	var sb strings.Builder
 	sb.WriteString("Given the user input you have the following in memory:\n")
-	for i, r := range searchResp.Results {
-		sb.WriteString(fmt.Sprintf("- %s", r.Content))
-		if len(r.Metadata) > 0 {
-			meta, _ := json.Marshal(r.Metadata)
-			sb.WriteString(fmt.Sprintf(" (%s)", string(meta)))
+
+	var citations []KBCitation
+	seen := make(map[string]struct{})
+
+	for _, r := range searchResp.Results {
+		fileName := r.Metadata[kbMetadataFileName]
+		source := r.Metadata[kbMetadataSource]
+
+		label := fileName
+		if label == "" {
+			label = "unknown"
 		}
-		if i < len(searchResp.Results)-1 {
-			sb.WriteString("\n")
+		sb.WriteString(fmt.Sprintf("[Source: %s]\n%s\n", label, r.Content))
+
+		// Citations are de-duplicated per source document: many chunks from the
+		// same file share one source key, so a file is listed only once. Skip
+		// results with no source key — they cannot be linked back to a document.
+		dedupKey := source
+		if dedupKey == "" {
+			dedupKey = fileName
 		}
+		if dedupKey == "" {
+			continue
+		}
+		if _, ok := seen[dedupKey]; ok {
+			continue
+		}
+		seen[dedupKey] = struct{}{}
+		citations = append(citations, KBCitation{
+			FileName: fileName,
+			EntryKey: source,
+		})
 	}
 
-	return sb.String()
+	sb.WriteString("When answering, cite sources using [Source: filename].")
+
+	return KBSearchContext{
+		Prompt:    sb.String(),
+		Citations: citations,
+	}
 }
 
 // KBSearchMemoryArgs defines the arguments for the search_memory tool.
@@ -110,21 +175,25 @@ type KBSearchMemoryArgs struct {
 
 // KBSearchMemoryTool implements the search_memory MCP tool.
 type KBSearchMemoryTool struct {
-	APIURL     string
-	APIKey     string
-	Collection string
-	MaxResults int
-	UserID     string
+	APIURL            string
+	APIKey            string
+	Collection        string
+	MaxResults        int
+	UserID            string
+	CitationCollector KBCitationCollector
 }
 
 func (t KBSearchMemoryTool) Run(args KBSearchMemoryArgs) (string, any, error) {
 	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
 	defer cancel()
 	result := KBAutoSearchPrompt(ctx, t.APIURL, t.APIKey, t.Collection, args.Query, t.MaxResults, t.UserID)
-	if result == "" {
+	if result.Prompt == "" {
 		return "No results found.", nil, nil
 	}
-	return result, nil, nil
+	if t.CitationCollector != nil {
+		t.CitationCollector.AddKBCitations(result.Citations)
+	}
+	return result.Prompt, nil, nil
 }
 
 // KBAddMemoryArgs defines the arguments for the add_memory tool.
@@ -156,9 +225,11 @@ func (t KBAddMemoryTool) Run(args KBAddMemoryArgs) (string, any, error) {
 
 // KBStoreContent uploads text content to a collection via the multipart upload API.
 func KBStoreContent(ctx context.Context, apiURL, apiKey, collection, content, userID string) error {
-	uploadURL := strings.TrimRight(apiURL, "/") + "/api/agents/collections/" + collection + "/upload"
+	uploadURL := strings.TrimRight(apiURL, "/") + "/api/agents/collections/" + url.PathEscape(collection) + "/upload"
 	if userID != "" {
-		uploadURL += "?user_id=" + userID
+		query := url.Values{}
+		query.Set("user_id", userID)
+		uploadURL += "?" + query.Encode()
 	}
 
 	// Build multipart form with the text content as a file