@tablestore/tablestore-agent-cli 0.2.3 → 0.2.4

package/dist/dashboard-runtime/dashboard/src/app/(dashboard)/memory/[store]/units/page.tsx

@@ -25,7 +25,11 @@ export default function MemoryUnitsPage() {
   const params = useParams<{ store: string }>();
   const searchParams = useSearchParams();
   const memoryStoreName = decodeURIComponent(params.store);
-  const scope = useMemo(() => buildMemoryScope(searchParams), [searchParams.toString()]);
+  const searchParamsString = searchParams.toString();
+  const scope = useMemo(
+    () => buildMemoryScope(new URLSearchParams(searchParamsString)),
+    [searchParamsString],
+  );
   const wildcard = hasWildcardScope(scope);
   const scopeKey = useMemo(() => 
     `${scope.appId}|${scope.tenantId ?? ""}|${scope.agentId ?? ""}|${scope.runId ?? ""}`,
@@ -87,7 +91,7 @@ export default function MemoryUnitsPage() {
       return;
     }
     if (!canOperateOnUnit(deleteUnit, scope)) {
-      toast.error("当前记忆单元缺少精确 scope，无法删除");
+      toast.error("当前记忆单元缺少精确范围，无法删除");
       return;
     }
     try {
@@ -102,11 +106,11 @@ export default function MemoryUnitsPage() {
 
   async function handleSearchById() {
     if (!searchId.trim()) {
-      toast.error("请输入 Memory ID");
+      toast.error("请输入记忆 ID");
       return;
     }
     if (!scope.appId) {
-      toast.error("请先设置精确的 Scope");
+      toast.error("请先设置精确的范围");
       return;
     }
     setSearching(true);
@@ -170,7 +174,7 @@ export default function MemoryUnitsPage() {
                   </span>
                 </TooltipTrigger>
                 <TooltipContent side="bottom" className="max-w-[240px]">
-                  <p>请在上方 Scope 栏中为所有字段设置具体值后再添加记忆</p>
+                  <p>请在上方范围栏中为所有字段设置具体值后再添加记忆</p>
                 </TooltipContent>
               </Tooltip>
             </TooltipProvider>
@@ -187,7 +191,7 @@ export default function MemoryUnitsPage() {
         <Input
           value={searchId}
           onChange={(e) => setSearchId(e.target.value)}
-          placeholder="输入 Memory ID 查询单条记忆..."
+          placeholder="输入记忆 ID 查询单条记忆..."
           className="max-w-md"
           onKeyDown={(e) => e.key === "Enter" && handleSearchById()}
         />
@@ -201,7 +205,7 @@ export default function MemoryUnitsPage() {
 
       {wildcard && !searchedUnit && (
         <p className="text-xs text-muted-foreground">
-          当前 scope 含通配符，仅支持只读查询，无法执行新增、查看详情、编辑或删除。
+          当前范围含通配符，仅支持只读查询，无法执行新增、查看详情、编辑或删除。
         </p>
       )}
 

package/dist/dashboard-runtime/dashboard/src/app/(dashboard)/memory/create/page.tsx

@@ -18,7 +18,7 @@ export default function CreateMemoryStorePage() {
         </Link>
         <PageHeader
           title="创建记忆库"
-          description="创建一个新的 Agent Memory store"
+          description="创建一个新的智能体记忆库"
         />
       </div>
 

package/dist/dashboard-runtime/dashboard/src/app/(dashboard)/memory/page.tsx

@@ -11,7 +11,7 @@ export default function MemoryPage() {
     <div className="space-y-6">
       <PageHeader
         title="记忆库"
-        description="管理 Agent 的长期记忆、检索记忆单元并观测请求记录"
+        description="管理智能体的长期记忆、检索记忆单元并观测请求记录"
       >
         <Button asChild>
           <Link href="/memory/create">

package/dist/dashboard-runtime/dashboard/src/app/api/v1/llm/chat/route.ts

@@ -1,6 +1,6 @@
 import { type NextRequest } from "next/server";
 
-import { chatCompletionStream } from "@/lib/server/llm-client";
+import { chatCompletionStream, type LLMMessageContent, type ContentPart } from "@/lib/server/llm-client";
 import { getProviderWithSecret } from "@/lib/server/llm-provider-store";
 import { rewriteQuery } from "@/lib/server/query-rewriter";
 import {
@@ -74,8 +74,24 @@ function buildDefaultPreamble(opts: {
 // Matches the plan's simple "keep last 20 turns" policy; token-based trimming
 // can land later when we wire up `tiktoken`.
 const MAX_HISTORY_MESSAGES = 20;
+/**
+ * 单个检索 chunk 拼入 system message 时的最大显示字符数。
+ * 用于控制 system 提示总长度，避免单个超长 chunk 把模型上下文挤爆。
+ */
+const MAX_CHUNK_DISPLAY_LENGTH = 1200;
 const chatLogger = createLogger("dashboard.api.llm");
 
+/**
+ * 判断一个检索 chunk 是否为图片类型。
+ *
+ * 服务端契约下 chunkType 共 4 种取值：text | table | formula | image。
+ * 大小写不敏感地比较，以兼容后端契约可能出现的大小写漂移。除 image 外的
+ * 3 种类型都按文本走 contextBlock 拼接，不进入多模态 image_url 部分。
+ */
+function isImageChunk(chunk: { chunkType?: string }): boolean {
+  return chunk.chunkType?.toUpperCase() === "IMAGE";
+}
+
 // turnId 去重：客户端每个「逻辑轮次」生成一次 turnId（regenerate / edit-resend
 // 复用同一个值）。我们仅在内存里维护一个 ~10 分钟过期的 set；不要求扛过进程重启。
 const TURN_DEDUP_TTL_MS = 10 * 60 * 1000;
@@ -533,13 +549,14 @@ async function resolveMemoryClient(request: NextRequest): Promise<
  * cite back. The header mentions the KB count when multiple KBs were queried
  * so the model can reason about source diversity.
  */
-function buildFinalMessages(
+/** @internal Exported for unit testing multimodal message construction. */
+export function buildFinalMessages(
   userSystemPrompt: string,
   chunks: RetrievedChunk[],
   groups: RetrievalGroup[],
   memoryBlock: string,
   history: Array<{ role: ChatRole; content: string }>,
-): Array<{ role: ChatRole; content: string }> {
+): Array<{ role: ChatRole; content: LLMMessageContent }> {
   const preamble =
     userSystemPrompt.trim() ||
     buildDefaultPreamble({
@@ -550,26 +567,120 @@ function buildFinalMessages(
   const header = successfulKbCount > 1
     ? `检索结果（共 ${chunks.length} 条，来自 ${successfulKbCount} 个知识库）：`
     : `检索结果（共 ${chunks.length} 条）：`;
+
+  // Strip any `system` messages the client sent — we own the system turn.
+  const nonSystem = history.filter((m) => m.role !== "system");
+  const clipped =
+    nonSystem.length > MAX_HISTORY_MESSAGES
+      ? nonSystem.slice(-MAX_HISTORY_MESSAGES)
+      : nonSystem;
+
+  // Collect image parts with numbered labels to preserve chunk order.
+  //
+  // ChunkType 服务端契约共 4 种：text | table | formula | image。
+  // 这 4 种里只有 image 需要走 `image_url` 部分注入到 user message；其他全部
+  // 作为文本，由下面的 `contextBlock` 直接以 chunk.content 形式拼进 system
+  // message。这里大小写不敏感地匹配 IMAGE，使后端 chunkType 字段大小写漂移
+  // 不会破坏多模态注入。
+  const imageParts: ContentPart[] = [];
+  chunks.forEach((c, i) => {
+    if (isImageChunk(c) && c.imageBase64) {
+      // Add numbered label so the model can correlate with system message [N].
+      // 编号 `i + 1` 与下方 contextBlock 中的 `[N]` 保持同源，确保模型能把
+      // 「检索图片 2」对应到 system 提示里的 `[2]` 引用。
+      imageParts.push({
+        type: "text",
+        text: `[检索图片 ${i + 1}]：`,
+      });
+      imageParts.push({
+        type: "image_url",
+        image_url: { url: c.imageBase64 },
+      });
+    }
+  });
+
+  // System message is always a plain string (works with all providers).
+  // Chunks 在 system 中按检索返回顺序展开，并保持与 imageParts 同一编号系统。
   const contextBlock = chunks.length
     ? `\n\n${header}\n${chunks
         .map((c, i) => {
           const src = c.knowledgeBaseName ? ` (${c.knowledgeBaseName})` : "";
-          return `[${i + 1}]${src} ${truncate(c.content, 1200)}`;
+          // 仅在 IMAGE 且没有 caption/content 时用 "(图片)" 占位；其余 chunk
+          // 类型（text / table / formula）一律直接展开 content。
+          const body = truncate(
+            c.content || (isImageChunk(c) ? "(图片)" : ""),
+            MAX_CHUNK_DISPLAY_LENGTH,
+          );
+          return `[${i + 1}]${src} ${body}`;
         })
         .join("\n\n")}`
     : "";
 
-  // Strip any `system` messages the client sent — we own the system turn.
-  const nonSystem = history.filter((m) => m.role !== "system");
-  const clipped =
-    nonSystem.length > MAX_HISTORY_MESSAGES
-      ? nonSystem.slice(-MAX_HISTORY_MESSAGES)
-      : nonSystem;
+  const systemMessage = {
+    role: "system" as const,
+    content: preamble + contextBlock + memoryBlock,
+  };
+
+  // If no images, return plain text messages (backward compatible)
+  if (imageParts.length === 0) {
+    return [systemMessage, ...clipped];
+  }
 
-  return [
-    { role: "system", content: preamble + contextBlock + memoryBlock },
-    ...clipped,
+  // Enrich the last user message with image parts
+  const messages: Array<{ role: ChatRole; content: LLMMessageContent }> = [
+    systemMessage,
   ];
+
+  const imageFallbackMessage = {
+    role: "user" as const,
+    content: [
+      ...imageParts,
+      { type: "text" as const, text: "请根据以上图片回答。" },
+    ],
+  };
+
+  if (clipped.length > 0) {
+    const lastIndex = clipped.length - 1;
+    const lastMessage = clipped[lastIndex];
+
+    if (lastMessage.role === "user") {
+      // Replace last user message with multimodal version
+      messages.push(
+        ...clipped.slice(0, lastIndex),
+        {
+          role: "user",
+          content: [
+            ...imageParts,
+            { type: "text", text: lastMessage.content },
+          ],
+        }
+      );
+    } else {
+      // Last message isn't a plain-text user message (e.g. assistant) — append
+      // a synthetic user turn carrying the images. This path is rare in
+      // production (clients normally end on a user turn) but can happen if the
+      // upstream injected a trailing assistant message or replayed history.
+      // Log a warning so we can spot it in observability rather than silently
+      // produce a "user said: 请根据以上图片回答。" turn the user never typed.
+      chatLogger.warn({
+        event: "api.chat.multimodal_fallback_user_turn",
+        message:
+          "last history message is not a user turn; appending synthetic " +
+          "user message to carry retrieved images",
+        details: {
+          lastRole: lastMessage.role,
+          imageCount: imageParts.length / 2,
+          historyLength: clipped.length,
+        },
+      });
+      messages.push(...clipped, imageFallbackMessage);
+    }
+  } else {
+    // No history — create a user message with images
+    messages.push(imageFallbackMessage);
+  }
+
+  return messages;
 }
 
 function truncate(text: string, max: number): string {

package/dist/dashboard-runtime/dashboard/src/app/globals.css

@@ -1,6 +1,25 @@
 @import "tailwindcss";
 @import "katex/dist/katex.min.css";
 
+/*
+ * Bridge the legacy Tailwind v3 style config (tailwind.config.ts) into
+ * Tailwind v4. Without @config, the JS config file is silently ignored,
+ * which means:
+ *   1. Theme tokens (success/warning/danger/info/purple) are never
+ *      registered, so utilities like `bg-success-soft` don't exist.
+ *   2. `darkMode: ["class", ".dark"]` is not applied, so `dark:` falls
+ *      back to `prefers-color-scheme`, which fights with next-themes'
+ *      class-based toggle and bleeds dark styles into light mode.
+ */
+@config "../../tailwind.config.ts";
+
+/*
+ * Belt-and-suspenders: ensure `dark:` always resolves through the
+ * `.dark` class that next-themes toggles on <html>, regardless of OS
+ * preference. This must agree with the darkMode setting above.
+ */
+@custom-variant dark (&:where(.dark, .dark *));
+
 @layer base {
   :root {
     /* Brand primary: webUI blue (hue 221, #2563eb) */
@@ -77,6 +96,8 @@
     --warning-soft: 48 100% 96%;
     --danger: 0 84% 60%;
     --danger-soft: 0 86% 97%;
+    --info: 217 91% 60%;
+    --info-soft: 214 100% 97%;
     --purple: 262 84% 58%;
     --purple-soft: 258 100% 97%;
 
@@ -174,6 +195,8 @@
     --warning-soft: 38 30% 18%;
     --danger: 0 84% 65%;
     --danger-soft: 0 30% 20%;
+    --info: 217 91% 70%;
+    --info-soft: 217 30% 18%;
     --purple: 262 84% 65%;
     --purple-soft: 262 30% 22%;
   }

package/dist/dashboard-runtime/dashboard/src/components/chat/chat-config-context.tsx

@@ -98,6 +98,7 @@ type Action =
   | { type: "SET_RETRIEVAL"; value: ChatRetrievalConfig }
   | { type: "ADD_MOUNT"; mount: MountItem }
   | { type: "UPDATE_MOUNT"; id: string; patch: Partial<MountItem> }
+  | { type: "RESET_MOUNT_RETRIEVAL"; id: string }
   | { type: "REMOVE_MOUNT"; id: string }
   | { type: "REORDER_MOUNTS"; ids: string[] }
   | { type: "SET_MEMORY"; binding: Omit<ChatMemoryBinding, "id"> | undefined }
@@ -136,6 +137,21 @@ export function reducer(state: ChatConfigState, action: Action): ChatConfigState
           m.id === action.id ? { ...m, ...action.patch } : m,
         ),
       };
+    // 清除单 KB 的自定义检索参数：透过 partial spread 调用者不能以 undefined
+    // 重置字段（{ retrieval: undefined } 仍会使原始键存在），这里重建 mount
+    // 折叠对象，确保字段从存在变为不存在，让后续 `??` fallback 生效。
+    case "RESET_MOUNT_RETRIEVAL":
+      return {
+        ...state,
+        mounts: state.mounts.map((m) => {
+          if (m.id !== action.id) return m;
+          // 按需保留 id / knowledgeBaseName / subspace；丢弃 retrieval / topK。
+          const { retrieval: _retrieval, topK: _topK, ...rest } = m;
+          void _retrieval;
+          void _topK;
+          return rest;
+        }),
+      };
     case "REMOVE_MOUNT":
       return { ...state, mounts: state.mounts.filter((m) => m.id !== action.id) };
     case "REORDER_MOUNTS": {
@@ -189,6 +205,8 @@ export interface ChatConfigContextValue extends ChatConfigState {
   setRetrieval(value: ChatRetrievalConfig): void;
   addMount(mount: MountItem): void;
   updateMount(id: string, patch: Partial<MountItem>): void;
+  /** Drop the per-KB retrieval/topK overrides for `id` so it inherits global defaults. */
+  resetMountRetrieval(id: string): void;
   removeMount(id: string): void;
   reorderMounts(ids: string[]): void;
   setMemory(binding: Omit<ChatMemoryBinding, "id"> | undefined): void;
@@ -217,6 +235,8 @@ export function ChatConfigProvider({ children }: { children: ReactNode }) {
       setRetrieval: (v) => dispatch({ type: "SET_RETRIEVAL", value: v }),
       addMount: (m) => dispatch({ type: "ADD_MOUNT", mount: m }),
       updateMount: (id, patch) => dispatch({ type: "UPDATE_MOUNT", id, patch }),
+      resetMountRetrieval: (id) =>
+        dispatch({ type: "RESET_MOUNT_RETRIEVAL", id }),
       removeMount: (id) => dispatch({ type: "REMOVE_MOUNT", id }),
       reorderMounts: (ids) => dispatch({ type: "REORDER_MOUNTS", ids }),
       setMemory: (b) => dispatch({ type: "SET_MEMORY", binding: b }),
@@ -256,27 +276,29 @@ export function useChatConfig(): ChatConfigContextValue {
  * return undefined when no KB is mounted. Kept as a plain helper so it can
  * run outside React rendering (e.g. in onClick handlers).
  *
- * The same `retrievalConfiguration` (searchType / rerank strategy / weights)
- * is applied to every mounted KB: cross-KB score comparability only holds
- * when each KB is scored under identical rules. `topK` is likewise uniform
- * per KB, and the caller (chat route) applies a global aggregate cap after
- * merging by score.
+ * 每个 mount 优先使用自己的 `retrieval` / `topK` override；未设置时 fallback
+ * 到会话级的 `state.retrieval` / `state.perKbTopK`。上层在 chat route 合并后
+ * 仍由 `state.aggregateTopK` 统一截断，跨 KB 的打分可比性责任交给服务端。
  */
 export function buildRagFromMounts(state: ChatConfigState) {
   if (!state.mounts.length) return undefined;
-  const retrievalConfiguration = assembleRetrievalConfiguration(
-    state.retrieval,
-    state.perKbTopK,
-  );
   return {
-    knowledgeBases: state.mounts.map((m) => ({
-      knowledgeBaseName: m.knowledgeBaseName,
-      ...(m.subspace?.length ? { subspace: m.subspace } : {}),
-      ...(retrievalConfiguration
-        ? { retrievalConfiguration: retrievalConfiguration as Record<string, unknown> }
-        : {}),
-      topK: state.perKbTopK,
-    })),
+    knowledgeBases: state.mounts.map((m) => {
+      const effectiveRetrieval = m.retrieval ?? state.retrieval;
+      const effectiveTopK = m.topK ?? state.perKbTopK;
+      const retrievalConfiguration = assembleRetrievalConfiguration(
+        effectiveRetrieval,
+        effectiveTopK,
+      );
+      return {
+        knowledgeBaseName: m.knowledgeBaseName,
+        ...(m.subspace?.length ? { subspace: m.subspace } : {}),
+        ...(retrievalConfiguration
+          ? { retrievalConfiguration: retrievalConfiguration as Record<string, unknown> }
+          : {}),
+        topK: effectiveTopK,
+      };
+    }),
     topK: state.aggregateTopK,
   };
 }

package/dist/dashboard-runtime/dashboard/src/components/chat/chat-mount-config.tsx

@@ -187,6 +187,7 @@ export function ChatMountConfig({ open, onOpenChange, onSave }: ChatMountConfigP
                         key={mount.id}
                         mount={mount}
                         subspaceEnabled={kb?.subspace === true}
+                        knowledgeBase={kb}
                       />
                     );
                   })}

package/dist/dashboard-runtime/dashboard/src/components/chat/memory-bind-section.tsx

@@ -45,8 +45,8 @@ export function MemoryBindSection() {
           <Label className="text-sm font-medium">绑定的记忆库</Label>
           <p className="mt-0.5 text-xs text-muted-foreground">
             {binding
-              ? `每轮检索 Top-${binding.topK ?? 5}，对话后写回当前 scope`
-              : "绑定后服务端在每轮对话前检索、对话后写回；同一 app+tenant 下记忆跨 agent 可共享"}
+              ? `每轮检索 Top-${binding.topK ?? 5}，对话后写回当前范围`
+              : "绑定后服务端在每轮对话前检索、对话后写回；同一应用 + 租户下记忆跨智能体可共享"}
           </p>
         </div>
         {binding && (
@@ -146,7 +146,7 @@ function MemoryBindEmptyState(props: {
       <p className="mx-auto mt-1 max-w-[320px] text-xs leading-relaxed text-[hsl(var(--muted-foreground))]">
         {hasNoStores
           ? "创建第一个记忆库后，即可在对话中绑定并自动写回 / 检索"
-          : "绑定后服务端会在每轮对话前检索相关记忆注入到 system prompt，对话结束后自动写回到当前 scope"}
+          : "绑定后服务端会在每轮对话前检索相关记忆注入到 system prompt，对话结束后自动写回到当前范围"}
       </p>
 
       <div className="mt-4 flex flex-col items-center gap-2">
@@ -247,27 +247,27 @@ function MemoryBindingForm(props: {
       <div className="space-y-2.5">
         <div>
           <Label className="text-xs font-medium uppercase tracking-wider text-muted-foreground">
-            写 Scope
+            写入范围
           </Label>
           <p className="mt-1 text-[11px] leading-relaxed text-muted-foreground">
-            服务端在检索时把 agentId / runId 派生为通配 <code className="rounded bg-muted px-1 py-0.5 font-mono text-[10px]">*</code>，
-            因此同一 app + tenant 下不同 agent / 不同会话可以读取彼此的记忆
+            服务端在检索时把「智能体 / 会话」派生为通配 <code className="rounded bg-muted px-1 py-0.5 font-mono text-[10px]">*</code>，
+            因此同一应用 + 租户下不同智能体 / 不同会话可以读取彼此的记忆
           </p>
         </div>
         <div className="grid grid-cols-2 gap-3">
           <ScopeField
-            label="appId *"
+            label="应用 *"
             value={props.appId}
             onChange={(v) => props.onScopeChange({ appId: v })}
           />
           <ScopeField
-            label="tenantId *"
+            label="租户 *"
             value={props.tenantId}
             onChange={(v) => props.onScopeChange({ tenantId: v })}
           />
           <div className="col-span-2">
             <ScopeField
-              label="agentId *"
+              label="智能体 *"
               value={props.agentId}
               onChange={(v) => props.onScopeChange({ agentId: v })}
             />
@@ -279,7 +279,7 @@ function MemoryBindingForm(props: {
         <div className="space-y-0.5">
           <Label className="text-sm">对话后自动写入记忆</Label>
           <p className="text-xs text-muted-foreground">
-            autoSave：把本轮 user / assistant 消息归档到当前 scope
+            自动归档：把本轮用户 / 助手消息写入当前范围
           </p>
         </div>
         <Switch
@@ -327,7 +327,7 @@ function RunIdField({ runId }: { runId: string }) {
     try {
       await navigator.clipboard.writeText(runId);
       setCopied(true);
-      toast.success("已复制 runId");
+      toast.success("已复制会话 ID");
       setTimeout(() => setCopied(false), 1500);
     } catch {
       toast.error("复制失败，请手动选择文本");
@@ -338,13 +338,13 @@ function RunIdField({ runId }: { runId: string }) {
     <div className="space-y-1.5">
       <div className="flex items-center justify-between">
         <Label className="text-xs font-medium uppercase tracking-wider text-muted-foreground">
-          run-id
+          会话 ID
         </Label>
         <span
           className="text-[10px] text-muted-foreground"
-          title="runId 由当前会话自动生成并在整个会话内复用，用于串联同一对话的记忆"
+          title="会话 ID 由当前对话自动生成并在整个对话内复用，用于串联同一会话的记忆"
         >
-          会话自动生成 · 自动复用
+          自动生成 · 自动复用
         </span>
       </div>
       <div className="flex items-center gap-2">
@@ -356,7 +356,7 @@ function RunIdField({ runId }: { runId: string }) {
           variant="ghost"
           size="icon"
           className="h-8 w-8 shrink-0"
-          aria-label="复制 runId"
+          aria-label="复制会话 ID"
           disabled={!runId}
           onClick={handleCopy}
         >