diff --git a/frontend-web/src/components/workspace/chats/chat-box.tsx b/frontend-web/src/components/workspace/chats/chat-box.tsx
index 8de0d29a..5687a78f 100644
--- a/frontend-web/src/components/workspace/chats/chat-box.tsx
+++ b/frontend-web/src/components/workspace/chats/chat-box.tsx
@@ -86,7 +86,11 @@ const ChatBox: React.FC<{
   threadId,
   disableArtifacts = false,
 }) => {
-  const { isMock, thread } = useThread();
+  const {
+    isMock,
+    thread,
+    knowledgeReferencesEnabled = false,
+  } = useThread();
   const pathname = usePathname();
   const threadIdRef = useRef(threadId);
   const artifactsCountRef = useRef(0);
@@ -107,7 +111,8 @@ const ChatBox: React.FC<{
     citationDisplay?.reference_mode_enabled ?? false;
   const ragModeEnabled =
     (referenceModeEntryEnabled && settings.display.ragModeEnabled) ||
-    isThreadRagModeLocked(threadId);
+    isThreadRagModeLocked(threadId) ||
+    knowledgeReferencesEnabled;
   // Frontend RAG modes write their own hidden source tool messages. They use
   // the existing reference panel, but must not request or write reference
   // batches through DeerFlow's backend APIs.
@@ -121,7 +126,7 @@ const ChatBox: React.FC<{
   );
   const referencesEnabled = ragModeEnabled || hasFrontendRag;
   const shouldFetchBackendReferenceBatches =
-    referencesEnabled && !hasFrontendRag;
+    referencesEnabled && !hasFrontendRag && !!threadId && threadId !== "new";
   const { skills } = useSkills();
 
   const canScanArtifactDirectory =
diff --git a/frontend-web/src/components/workspace/messages/context.ts b/frontend-web/src/components/workspace/messages/context.ts
index d22189c7..65f3a626 100644
--- a/frontend-web/src/components/workspace/messages/context.ts
+++ b/frontend-web/src/components/workspace/messages/context.ts
@@ -4,11 +4,17 @@ import { createContext, useContext } from "react";
 import type { RuntimeDisplaySource } from "@/core/skill-display/runtime";
 import type { AgentThreadState } from "@/core/threads";
 
-export interface ThreadContextType {
-  thread: BaseStream<AgentThreadState>;
-  isMock?: boolean;
-  /**
-   * When set, the conversation is rendered from a public read-only share.
+export interface ThreadContextType {
+  thread: BaseStream<AgentThreadState>;
+  isMock?: boolean;
+  /**
+   * True when the active conversation has selected LLMWiki knowledge bases.
+   * Knowledge-base Q&A should always expose its source documents, independent
+   * from the admin-controlled manual reference-mode button.
+   */
+  knowledgeReferencesEnabled?: boolean;
+  /**
+   * When set, the conversation is rendered from a public read-only share.
    * Artifact/image URLs then resolve to the unauthenticated
    * `/api/public/shares/{code}/files/...` endpoint instead of the
    * auth-gated thread artifact route.
diff --git a/frontend-web/src/components/workspace/messages/message-list.tsx b/frontend-web/src/components/workspace/messages/message-list.tsx
index 15d83aff..e5cae494 100644
--- a/frontend-web/src/components/workspace/messages/message-list.tsx
+++ b/frontend-web/src/components/workspace/messages/message-list.tsx
@@ -305,9 +305,12 @@ export function MessageList({
   const { data: citationDisplay } = useCitationDisplaySettings();
   const referenceModeEntryEnabled =
     citationDisplay?.reference_mode_enabled ?? false;
+  const knowledgeReferencesEnabled =
+    threadContext?.knowledgeReferencesEnabled === true;
   const ragModeEnabled =
     (referenceModeEntryEnabled && settings.display.ragModeEnabled) ||
-    isThreadRagModeLocked(threadId);
+    isThreadRagModeLocked(threadId) ||
+    knowledgeReferencesEnabled;
   const hasSingleRoundRag = useMemo(
     () =>
       messages.some(
diff --git a/frontend-web/src/pages/AgentChatPage.tsx b/frontend-web/src/pages/AgentChatPage.tsx
index 20774b4f..80f9595b 100644
--- a/frontend-web/src/pages/AgentChatPage.tsx
+++ b/frontend-web/src/pages/AgentChatPage.tsx
@@ -779,7 +779,13 @@ export default function AgentChatPage({
       : "ready";
 
   return (
-    <ThreadContext.Provider value={{ thread: displayThread }}>
+    <ThreadContext.Provider
+      value={{
+        thread: displayThread,
+        knowledgeReferencesEnabled:
+          knowledgeSelectionReady && selectedKnowledgeBaseIds.length > 0,
+      }}
+    >
       {/* Root: fixed-height horizontal split. ChatBox (chat + artifact
           preview) lives in the left flex cell; the agent detail panel is a
           separate flex sibling OUTSIDE ChatBox, so the artifact preview
diff --git a/frontend-web/src/pages/ChatPage.tsx b/frontend-web/src/pages/ChatPage.tsx
index c52559cf..2a7a93a7 100644
--- a/frontend-web/src/pages/ChatPage.tsx
+++ b/frontend-web/src/pages/ChatPage.tsx
@@ -688,7 +688,14 @@ export default function ChatPage() {
     !isWujieEnv();
 
   return (
-    <ThreadContext.Provider value={{ thread: displayThread, isMock }}>
+    <ThreadContext.Provider
+      value={{
+        thread: displayThread,
+        isMock,
+        knowledgeReferencesEnabled:
+          knowledgeSelectionReady && selectedKnowledgeBaseIds.length > 0,
+      }}
+    >
       <ChatBox threadId={threadId}>
         <div
           className={cn(
diff --git a/offline-backend-20260512/backend/CLAUDE.md b/offline-backend-20260512/backend/CLAUDE.md
index de8b34c5..93cd1461 100644
--- a/offline-backend-20260512/backend/CLAUDE.md
+++ b/offline-backend-20260512/backend/CLAUDE.md
@@ -486,7 +486,8 @@ frontend can hide or show the entry consistently; `PUT` is admin-only and
 partially updates `reference_mode_enabled` and/or `cleanup_words`. The normal
 workspace input box, iframe chat toolbar, send context, right reference panel,
 and appearance settings all use this flag, while already locked citation
-threads and knowledge-base selections still render their references. Tests:
+threads and selected LLMWiki knowledge bases still render their source
+documents in the reference panel. Tests:
 `tests/test_prompt_prefix.py` and
 `frontend-web/src/pages/iframe-chat-mode-buttons.test.ts`.
 
diff --git a/offline-backend-20260512/backend/README.md b/offline-backend-20260512/backend/README.md
index efef9706..584ce987 100644
--- a/offline-backend-20260512/backend/README.md
+++ b/offline-backend-20260512/backend/README.md
@@ -670,7 +670,9 @@ Reference-mode entry points are controlled by `citation_display` in
 `.deer-flow/system_settings.json`. `reference_mode_enabled` defaults to `false`;
 when disabled, the frontend hides the "参考文献" composer button in both the
 normal workspace and iframe chat, and hides the per-user RAG citation toggle
-from appearance settings. `GET /api/system-settings/citation-display` is
+from appearance settings. Conversations with selected LLMWiki knowledge bases
+still show their source documents in the reference panel even when the manual
+entry is hidden. `GET /api/system-settings/citation-display` is
 readable by logged-in users so the UI can decide whether to expose the entry;
 `PUT /api/system-settings/citation-display` is admin-only and supports partial
 updates for both `reference_mode_enabled` and `cleanup_words`.