feat(memory): hot-swap embedding provider on rebuild-index

Switching embedding provider in the web console no longer requires a
restart and no longer drops the running conversation

channel/web/static/js/console.js

@@ -50,15 +50,20 @@ const I18N = {
         models_unavailable: '不可用',
         models_set_via_env: '通过环境变量启用',
         models_dim_label: '维度',
-        models_warn_rebuild: '向量配置已更新，建议在聊天框输入 /memory rebuild-index 重建索引',
         models_save_success: '已保存',
         models_save_failed: '保存失败',
         models_cleared: '已清除',
         models_clear_failed: '清除失败',
+        models_embedding_change_title: '更改向量模型',
+        models_embedding_change_msg: '切换向量模型后，已有索引将失效，需要重建。是否继续？',
+        models_embedding_saved_title: '向量模型已更新',
+        models_embedding_saved_msg: '请在聊天框输入 /memory rebuild-index 重建索引。',
+        models_embedding_saved_ok: '去执行',
         models_clear_confirm_title: '清除厂商凭据',
         models_clear_confirm_msg: '确认清除该厂商的 API Key 与 Base URL 吗？相关能力将不再可用。',
         cancel: '取消',
         save: '保存',
+        ok: '确定',
         knowledge_title: '知识库', knowledge_desc: '浏览和探索你的知识库',
         knowledge_tab_docs: '文档', knowledge_tab_graph: '图谱',
         knowledge_loading: '加载知识库中...', knowledge_loading_desc: '知识页面将显示在这里',
@@ -193,15 +198,20 @@ const I18N = {
         models_unavailable: 'unavailable',
         models_set_via_env: 'enable via environment variable',
         models_dim_label: 'dim',
-        models_warn_rebuild: 'Embedding settings changed — run /memory rebuild-index in chat to rebuild',
         models_save_success: 'Saved',
         models_save_failed: 'Save failed',
         models_cleared: 'Cleared',
         models_clear_failed: 'Clear failed',
+        models_embedding_change_title: 'Change embedding model',
+        models_embedding_change_msg: 'Switching the embedding model invalidates the existing index — a rebuild will be needed. Continue?',
+        models_embedding_saved_title: 'Embedding model updated',
+        models_embedding_saved_msg: 'Send /memory rebuild-index in the chat to rebuild the index.',
+        models_embedding_saved_ok: 'Go',
         models_clear_confirm_title: 'Clear vendor credentials',
         models_clear_confirm_msg: 'Remove this vendor\'s API Key and Base URL? Capabilities relying on it will stop working.',
         cancel: 'Cancel',
         save: 'Save',
+        ok: 'OK',
         knowledge_title: 'Knowledge', knowledge_desc: 'Browse and explore your knowledge base',
         knowledge_tab_docs: 'Documents', knowledge_tab_graph: 'Graph',
         knowledge_loading: 'Loading knowledge base...', knowledge_loading_desc: 'Knowledge pages will be displayed here',
@@ -3213,12 +3223,14 @@ function closeMemoryViewer() {
 // =====================================================================
 // Custom Confirm Dialog
 // =====================================================================
-function showConfirmDialog({ title, message, okText, cancelText, onConfirm }) {
+function showConfirmDialog({ title, message, okText, cancelText, onConfirm, hideCancel }) {
     const overlay = document.getElementById('confirm-dialog-overlay');
     document.getElementById('confirm-dialog-title').textContent = title || '';
     document.getElementById('confirm-dialog-message').textContent = message || '';
     document.getElementById('confirm-dialog-ok').textContent = okText || 'OK';
-    document.getElementById('confirm-dialog-cancel').textContent = cancelText || t('channels_cancel');
+    const cancelBtn = document.getElementById('confirm-dialog-cancel');
+    cancelBtn.textContent = cancelText || t('channels_cancel');
+    cancelBtn.classList.toggle('hidden', !!hideCancel);
 
     function cleanup() {
         overlay.classList.add('hidden');
@@ -3231,7 +3243,6 @@ function showConfirmDialog({ title, message, okText, cancelText, onConfirm }) {
     function onOverlayClick(e) { if (e.target === overlay) cleanup(); }
 
     const okBtn = document.getElementById('confirm-dialog-ok');
-    const cancelBtn = document.getElementById('confirm-dialog-cancel');
     okBtn.addEventListener('click', onOk);
     cancelBtn.addEventListener('click', onCancel);
     overlay.addEventListener('click', onOverlayClick);
@@ -3555,11 +3566,16 @@ function renderCapabilityBody(def, cap, body) {
     // For auto-capable capabilities, an "auto" strategy means the user has
     // not pinned a vendor; we honor that by selecting the empty-string
     // sentinel rather than the resolved fallback provider name.
+    // `suggested_provider` is a UI-only preselect for embedding when nothing
+    // is pinned yet — purely cosmetic, not persisted until the user saves.
     const initialProviderValue = pendingProvider
         ? pendingProvider
         : ((cap.strategy === 'auto' && capabilitySupportsAuto(def.id))
             ? ''
-            : (cap.current_provider || (ddOpts[0] && ddOpts[0].value) || ''));
+            : (cap.current_provider
+                || cap.suggested_provider
+                || (ddOpts[0] && ddOpts[0].value)
+                || ''));
     initDropdown(
         provDd,
         ddOpts,
@@ -3680,7 +3696,9 @@ function buildCapabilityProviderOptions(def, cap) {
 }
 
 function capabilitySupportsAuto(capId) {
-    return capId === 'image' || capId === 'vision' || capId === 'embedding';
+    // Embedding is intentionally NOT here: runtime only auto-falls back to
+    // OpenAI/LinkAI, so dressing it up as "auto" hides reality from users.
+    return capId === 'image' || capId === 'vision';
 }
 
 // After initDropdown renders the capability provider menu, decorate each
@@ -3882,21 +3900,64 @@ function saveCapability(capId) {
     const isAuto = provider === '' && capabilitySupportsAuto(capId);
     const model = isAuto ? '' : getCapabilityModelValue(def);
 
+    // Embedding changes invalidate any pre-existing vector index because
+    // dimensions / vendor differ. Gate the save behind a confirm, and on
+    // success surface a dedicated info dialog telling the user how to
+    // rebuild — both via the in-app custom dialog, not the native alert.
+    if (capId === 'embedding') {
+        const cap = modelsState.capabilities[capId] || {};
+        const before = (cap.current_provider || '').trim();
+        const after = (provider || '').trim();
+        if (before !== after) {
+            showConfirmDialog({
+                title: t('models_embedding_change_title'),
+                message: t('models_embedding_change_msg'),
+                okText: t('save'),
+                cancelText: t('cancel'),
+                onConfirm: () => _persistCapability(capId, provider, model, () => {
+                    showConfirmDialog({
+                        title: t('models_embedding_saved_title'),
+                        message: t('models_embedding_saved_msg'),
+                        okText: t('models_embedding_saved_ok'),
+                        hideCancel: true,
+                        onConfirm: () => {
+                            navigateTo('chat');
+                            // Defer focus + value set: navigateTo may
+                            // re-render the chat panel; setting value before
+                            // the input is mounted would be lost.
+                            setTimeout(() => {
+                                const input = document.getElementById('chat-input');
+                                if (!input) return;
+                                input.value = '/memory rebuild-index';
+                                input.focus();
+                                // Trigger any input listeners (autosize, send-button enable, etc.)
+                                input.dispatchEvent(new Event('input', { bubbles: true }));
+                            }, 60);
+                        },
+                    });
+                }),
+            });
+            return;
+        }
+    }
+    _persistCapability(capId, provider, model);
+}
+
+function _persistCapability(capId, provider, model, onAfterSuccess) {
     fetch('/api/models', {
         method: 'POST',
         headers: { 'Content-Type': 'application/json' },
         body: JSON.stringify({ action: 'set_capability', capability: capId, provider_id: provider, model: model }),
     }).then(r => r.json()).then(data => {
         if (data.status === 'success') {
-            // Show "Saved" first; defer the view refresh so the user can
-            // see the confirmation (loadModelsView rebuilds the card and
-            // would otherwise wipe the status span instantly).
+            // Show "Saved" first, then refresh — loadModelsView would
+            // otherwise rebuild the card and wipe the status span before
+            // the user can register the confirmation.
             showStatus(`cap-${capId}-status`, 'models_save_success', false);
             setTimeout(() => {
-                if (data.warn_rebuild_index) alert(t('models_warn_rebuild'));
-                // Re-pull state so derived fields (e.g. image follows main) refresh.
                 loadModelsView({ preserveScroll: true });
-            }, 900);
+                if (onAfterSuccess) onAfterSuccess();
+            }, 400);
         } else {
             showStatus(`cap-${capId}-status`, 'models_save_failed', true);
         }
@@ -4064,26 +4125,16 @@ function saveVendorModal() {
         headers: { 'Content-Type': 'application/json' },
         body: JSON.stringify(payload),
     }).then(r => r.json()).then(data => {
+        btn.disabled = false;
         if (data.status === 'success') {
-            // Show inline confirmation in the modal footer, then close after
-            // a short delay so the user can see the "Saved" feedback.
-            showStatus('vendor-modal-status', 'models_save_success', false);
-            setTimeout(() => {
-                btn.disabled = false;
-                closeVendorModal();
-                const onSaved = vendorModalState.onSaved;
-                // If a caller provided an onSaved hook, it owns the refresh
-                // strategy (e.g. preserveScroll when launched from inside a
-                // capability dropdown). Otherwise fall back to a plain reload
-                // so the standalone "add vendor" entry still updates the grid.
-                if (onSaved) {
-                    try { onSaved(providerId); } catch (e) { /* noop */ }
-                } else {
-                    loadModelsView();
-                }
-            }, 700);
+            closeVendorModal();
+            const onSaved = vendorModalState.onSaved;
+            if (onSaved) {
+                try { onSaved(providerId); } catch (e) { /* noop */ }
+            } else {
+                loadModelsView();
+            }
         } else {
-            btn.disabled = false;
             showStatus('vendor-modal-status', 'models_save_failed', true);
         }
     }).catch(() => {
@@ -4107,11 +4158,8 @@ function clearVendorModal() {
                 body: JSON.stringify({ action: 'delete_provider', provider_id: providerId }),
             }).then(r => r.json()).then(data => {
                 if (data.status === 'success') {
-                    showStatus('vendor-modal-status', 'models_cleared', false);
-                    setTimeout(() => {
-                        closeVendorModal();
-                        loadModelsView();
-                    }, 700);
+                    closeVendorModal();
+                    loadModelsView();
                 } else {
                     showStatus('vendor-modal-status', 'models_clear_failed', true);
                 }

channel/web/web_channel.py

@@ -1234,7 +1234,7 @@ class ModelsHandler:
     # ids drawn from ConfigHandler.PROVIDER_MODELS where applicable.
     _ASR_PROVIDERS = ["openai", "linkai", "baidu", "ali", "xunfei", "azure", "google"]
     _TTS_PROVIDERS = ["openai", "linkai", "minimax", "baidu", "ali", "xunfei", "azure", "google", "elevenlabs", "edge", "pytts"]
-    _EMBEDDING_PROVIDERS = ["openai", "linkai", "dashscope", "doubao", "zhipu"]
+    _EMBEDDING_PROVIDERS = ["openai", "dashscope", "doubao", "zhipu", "linkai"]
 
     # Capability-scoped model catalogs. The chat dropdown can reuse the
     # provider's generic model list, but vision and image generation are
@@ -1522,29 +1522,24 @@ class ModelsHandler:
 
     @classmethod
     def _embedding_capability(cls, local_config: dict) -> dict:
+        # Embedding is "pick or empty" — runtime's legacy openai/linkai
+        # fallback is a safety net, not a UX-visible auto mode.
+        # `suggested_provider` is a UI-only hint (NOT persisted) that
+        # preselects the dropdown to whichever configured vendor we'd
+        # recommend, so users don't have to expand the menu to find it.
         explicit = (local_config.get("embedding_provider") or "").strip().lower()
-        # When unset, the legacy auto path in agent_initializer.py picks
-        # openai -> linkai. We surface "auto" + an estimate of what it'd
-        # actually use, but don't probe the runtime here.
+        suggested = ""
         if not explicit:
-            if cls._is_real_key(local_config.get("open_ai_api_key", "")):
-                effective = "openai"
-            elif cls._is_real_key(local_config.get("linkai_api_key", "")):
-                effective = "linkai"
-            else:
-                effective = ""
-            return {
-                "editable": True,
-                "strategy": "auto",
-                "current_provider": effective,
-                "current_model": local_config.get("embedding_model", "") or "",
-                "current_dim": int(local_config.get("embedding_dimensions") or 0) or None,
-                "providers": cls._EMBEDDING_PROVIDERS,
-            }
+            for pid in cls._EMBEDDING_PROVIDERS:
+                meta = ConfigHandler.PROVIDER_MODELS.get(pid) or {}
+                key_field = meta.get("api_key_field")
+                if key_field and cls._is_real_key(local_config.get(key_field, "")):
+                    suggested = pid
+                    break
         return {
             "editable": True,
-            "strategy": "specified",
             "current_provider": explicit,
+            "suggested_provider": suggested,
             "current_model": local_config.get("embedding_model", "") or "",
             "current_dim": int(local_config.get("embedding_dimensions") or 0) or None,
             "providers": cls._EMBEDDING_PROVIDERS,
@@ -1931,15 +1926,10 @@ class ModelsHandler:
             file_cfg["embedding_model"] = ""
         self._write_file_config(file_cfg)
         logger.info(f"[ModelsHandler] embedding updated: provider={provider_id!r} model={model!r}")
-        # Embedding switches don't go through Bridge bots; the agent's
-        # MemoryManager rebuilds its provider on next process restart, but
-        # the index dim may now mismatch — frontend should warn the user.
-        return json.dumps({
-            "status": "success",
-            "provider": provider_id,
-            "model": model,
-            "warn_rebuild_index": True,
-        })
+        # The agent's MemoryManager picks the new provider on next process
+        # restart; the index dim may now mismatch so a rebuild is needed.
+        # The frontend surfaces this via a confirm + post-save dialog.
+        return json.dumps({"status": "success", "provider": provider_id, "model": model})
 
     @staticmethod
     def _reset_bridge() -> None:

plugins/cow_cli/cow_cli.py

@@ -1056,6 +1056,38 @@ class CowCliPlugin(Plugin):
             logger.warning(f"[CowCli] /memory dream sync failed: {e}")
             return f"❌ 记忆蒸馏失败: {e}"
 
+    @staticmethod
+    def _resolve_active_embedding():
+        """
+        Resolve (provider_label, model, dim) from the LATEST config, not the
+        possibly-stale provider instance cached on a running agent. Used by
+        /memory status and rebuild-index hints so they reflect what a rebuild
+        will actually run as after the user changes embedding_provider.
+        Returns (label, model, dim) where any field may be None when unknown.
+        """
+        from agent.memory.embedding import EMBEDDING_VENDORS
+        from config import conf
+
+        provider_key = (conf().get("embedding_provider") or "").strip().lower()
+        cfg_model = (conf().get("embedding_model") or "").strip()
+        try:
+            cfg_dim = int(conf().get("embedding_dimensions") or 0)
+        except (TypeError, ValueError):
+            cfg_dim = 0
+
+        if not provider_key:
+            # Legacy auto path: openai -> linkai, both default to text-embedding-3-small (1536).
+            if (conf().get("open_ai_api_key") or "").strip():
+                return "openai (legacy)", "text-embedding-3-small", 1536
+            if (conf().get("linkai_api_key") or "").strip():
+                return "linkai (legacy)", "text-embedding-3-small", 1536
+            return "(legacy)", None, None
+
+        meta = EMBEDDING_VENDORS.get(provider_key) or {}
+        model = cfg_model or meta.get("default_model")
+        dim = cfg_dim if cfg_dim > 0 else meta.get("default_dimensions")
+        return provider_key, model, dim
+
     def _memory_status(self) -> str:
         """Show current memory index status."""
         from agent.memory.embedding import detect_index_dim
@@ -1078,15 +1110,14 @@ class CowCliPlugin(Plugin):
         lines.append(f"  Chunks  : {chunks} (embedded: {embedded})")
         lines.append("")
 
-        # Active provider (from running config + provider instance).
+        # Resolve from the latest config so users see what /memory rebuild-index
+        # will actually run as — not what the cached agent was initialized with.
+        cfg_provider, cfg_model, cfg_dim = self._resolve_active_embedding()
         provider_obj = memory_manager.embedding_provider
-        cfg_provider = (conf().get("embedding_provider") or "").strip().lower() or "(legacy)"
-        if provider_obj is not None:
-            cfg_model = getattr(provider_obj, "model", "?")
-            cfg_dim = getattr(provider_obj, "_dimensions", None) or "?"
+        if cfg_model:
             lines.append(f"  Provider : {cfg_provider}")
             lines.append(f"  Model    : {cfg_model}")
-            lines.append(f"  Dim      : {cfg_dim}")
+            lines.append(f"  Dim      : {cfg_dim if cfg_dim else '?'}")
         else:
             lines.append("  Provider : (未初始化, keyword-only)")
 
@@ -1105,7 +1136,6 @@ class CowCliPlugin(Plugin):
                 )
 
             index_dim = detect_index_dim(memory_manager.storage)
-            cfg_dim = getattr(provider_obj, "_dimensions", None)
             if index_dim is not None and cfg_dim and index_dim != cfg_dim:
                 warnings.append(
                     f"  ⚠️ 索引中存量向量为 {index_dim} 维，与当前配置 {cfg_dim} 维不一致；"
@@ -1129,15 +1159,27 @@ class CowCliPlugin(Plugin):
             )
 
         memory_manager = agent.memory_manager
-        if memory_manager.embedding_provider is None:
+
+        # Rebuild against the LATEST config: build a fresh provider from
+        # config.json and swap it onto memory_manager. The agent's
+        # conversation_history and other state are untouched.
+        try:
+            from bridge.agent_initializer import AgentInitializer
+            fresh_provider = AgentInitializer(bridge=None, agent_bridge=None) \
+                ._init_embedding_provider(memory_manager.config, session_id=session_id)
+        except Exception as e:
+            logger.exception("[CowCli] /memory rebuild-index: build provider failed")
+            return f"⚠️ 无法根据当前配置构造 embedding provider: {e}"
+
+        if fresh_provider is None:
             return (
                 "⚠️ 当前没有可用的 embedding provider。\n"
                 "请检查 config.json 中的 embedding 相关配置 (provider / api key)。"
             )
+        memory_manager.embedding_provider = fresh_provider
 
-        provider_obj = memory_manager.embedding_provider
-        model_label = getattr(provider_obj, "model", "?")
-        dim_label = getattr(provider_obj, "dimensions", "?")
+        model_label = getattr(fresh_provider, "model", "?")
+        dim_label = getattr(fresh_provider, "dimensions", "?")
 
         # SaaS (e_context is None): run synchronously, return final result
         if e_context is None:
@@ -1168,7 +1210,7 @@ class CowCliPlugin(Plugin):
         threading.Thread(target=_run, daemon=True).start()
         return (
             f"🔧 索引重建已启动 (model={model_label}, dim={dim_label})\n\n"
-            f"将清空现有 chunks 并重新 embed 所有记忆文件，完成后会通知你。"
+            f"将重新向量化所有记忆和知识文件，完成后会通知你。"
         )
 
     @staticmethod