Merge b3aafba30c35c29b4126c2421e4ddde8f16e05d7 into 186249149674df5938faecabb3a3b7c48d9bce71

docs: add ClawCloud Run button

README.md

@@ -159,6 +159,7 @@ Perplexica runs on Next.js and handles all API requests. It works right away on
 
 [![Deploy to Sealos](https://raw.githubusercontent.com/labring-actions/templates/main/Deploy-on-Sealos.svg)](https://usw.sealos.io/?openapp=system-template%3FtemplateName%3Dperplexica)
 [![Deploy to RepoCloud](https://d16t0pc4846x52.cloudfront.net/deploylobe.svg)](https://repocloud.io/details/?app_id=267)
+[![Run on ClawCloud](https://raw.githubusercontent.com/ClawCloud/Run-Template/refs/heads/main/Run-on-ClawCloud.svg)](https://template.run.claw.cloud/?referralCode=U11MRQ8U9RM4&openapp=system-fastdeploy%3FtemplateName%3Dperplexica)
 
 ## Upcoming Features
 

package.json

@@ -20,6 +20,7 @@
     "@langchain/core": "^0.3.42",
     "@langchain/google-genai": "^0.1.12",
     "@langchain/openai": "^0.0.25",
+    "@langchain/ollama": "^0.2.0",
     "@langchain/textsplitters": "^0.1.0",
     "@tailwindcss/typography": "^0.5.12",
     "@xenova/transformers": "^2.17.2",

sample.config.toml

@@ -25,5 +25,8 @@ API_URL = "" # Ollama API URL - http://host.docker.internal:11434
 [MODELS.DEEPSEEK]
 API_KEY = ""
 
+[MODELS.LM_STUDIO]
+API_URL = "" # LM Studio API URL - http://host.docker.internal:1234
+
 [API_ENDPOINTS]
-SEARXNG = "" # SearxNG API URL - http://localhost:32768
+SEARXNG = "" # SearxNG API URL - http://localhost:32768

src/app/api/chat/route.ts

@@ -20,6 +20,7 @@ import {
   getCustomOpenaiApiUrl,
   getCustomOpenaiModelName,
 } from '@/lib/config';
+import { ChatOllama } from '@langchain/ollama';
 import { searchHandlers } from '@/lib/search';
 
 export const runtime = 'nodejs';
@@ -34,6 +35,7 @@ type Message = {
 type ChatModel = {
   provider: string;
   name: string;
+  ollamaContextWindow?: number;
 };
 
 type EmbeddingModel = {
@@ -232,6 +234,11 @@ export const POST = async (req: Request) => {
       }) as unknown as BaseChatModel;
     } else if (chatModelProvider && chatModel) {
       llm = chatModel.model;
+
+      // Set context window size for Ollama models
+      if (llm instanceof ChatOllama && body.chatModel?.provider === 'ollama') {
+        llm.numCtx = body.chatModel.ollamaContextWindow || 2048;
+      }
     }
 
     if (!llm) {

src/app/api/config/route.ts

@@ -8,6 +8,7 @@ import {
   getOllamaApiEndpoint,
   getOpenaiApiKey,
   getDeepseekApiKey,
+  getLMStudioApiEndpoint,
   updateConfig,
 } from '@/lib/config';
 import {
@@ -51,6 +52,7 @@ export const GET = async (req: Request) => {
 
     config['openaiApiKey'] = getOpenaiApiKey();
     config['ollamaApiUrl'] = getOllamaApiEndpoint();
+    config['lmStudioApiUrl'] = getLMStudioApiEndpoint();
     config['anthropicApiKey'] = getAnthropicApiKey();
     config['groqApiKey'] = getGroqApiKey();
     config['geminiApiKey'] = getGeminiApiKey();
@@ -93,6 +95,9 @@ export const POST = async (req: Request) => {
         DEEPSEEK: {
           API_KEY: config.deepseekApiKey,
         },
+        LM_STUDIO: {
+          API_URL: config.lmStudioApiUrl,
+        },
         CUSTOM_OPENAI: {
           API_URL: config.customOpenaiApiUrl,
           API_KEY: config.customOpenaiApiKey,

src/app/api/images/route.ts

@@ -7,11 +7,13 @@ import {
 import { getAvailableChatModelProviders } from '@/lib/providers';
 import { BaseChatModel } from '@langchain/core/language_models/chat_models';
 import { AIMessage, BaseMessage, HumanMessage } from '@langchain/core/messages';
+import { ChatOllama } from '@langchain/ollama';
 import { ChatOpenAI } from '@langchain/openai';
 
 interface ChatModel {
   provider: string;
   model: string;
+  ollamaContextWindow?: number;
 }
 
 interface ImageSearchBody {
@@ -58,6 +60,10 @@ export const POST = async (req: Request) => {
       }) as unknown as BaseChatModel;
     } else if (chatModelProvider && chatModel) {
       llm = chatModel.model;
+      // Set context window size for Ollama models
+      if (llm instanceof ChatOllama && body.chatModel?.provider === 'ollama') {
+        llm.numCtx = body.chatModel.ollamaContextWindow || 2048;
+      }
     }
 
     if (!llm) {

src/app/api/search/route.ts

@@ -13,12 +13,14 @@ import {
   getCustomOpenaiModelName,
 } from '@/lib/config';
 import { searchHandlers } from '@/lib/search';
+import { ChatOllama } from '@langchain/ollama';
 
 interface chatModel {
   provider: string;
   name: string;
   customOpenAIKey?: string;
   customOpenAIBaseURL?: string;
+  ollamaContextWindow?: number;
 }
 
 interface embeddingModel {
@@ -97,6 +99,10 @@ export const POST = async (req: Request) => {
         .model as unknown as BaseChatModel | undefined;
     }
 
+    if (llm instanceof ChatOllama && body.chatModel?.provider === 'ollama') {
+      llm.numCtx = body.chatModel.ollamaContextWindow || 2048;
+    }
+
     if (
       embeddingModelProviders[embeddingModelProvider] &&
       embeddingModelProviders[embeddingModelProvider][embeddingModel]

src/app/api/suggestions/route.ts

@@ -8,10 +8,12 @@ import { getAvailableChatModelProviders } from '@/lib/providers';
 import { BaseChatModel } from '@langchain/core/language_models/chat_models';
 import { AIMessage, BaseMessage, HumanMessage } from '@langchain/core/messages';
 import { ChatOpenAI } from '@langchain/openai';
+import { ChatOllama } from '@langchain/ollama';
 
 interface ChatModel {
   provider: string;
   model: string;
+  ollamaContextWindow?: number;
 }
 
 interface SuggestionsGenerationBody {
@@ -57,6 +59,10 @@ export const POST = async (req: Request) => {
       }) as unknown as BaseChatModel;
     } else if (chatModelProvider && chatModel) {
       llm = chatModel.model;
+      // Set context window size for Ollama models
+      if (llm instanceof ChatOllama && body.chatModel?.provider === 'ollama') {
+        llm.numCtx = body.chatModel.ollamaContextWindow || 2048;
+      }
     }
 
     if (!llm) {

src/app/api/videos/route.ts

@@ -7,11 +7,13 @@ import {
 import { getAvailableChatModelProviders } from '@/lib/providers';
 import { BaseChatModel } from '@langchain/core/language_models/chat_models';
 import { AIMessage, BaseMessage, HumanMessage } from '@langchain/core/messages';
+import { ChatOllama } from '@langchain/ollama';
 import { ChatOpenAI } from '@langchain/openai';
 
 interface ChatModel {
   provider: string;
   model: string;
+  ollamaContextWindow?: number;
 }
 
 interface VideoSearchBody {
@@ -58,6 +60,10 @@ export const POST = async (req: Request) => {
       }) as unknown as BaseChatModel;
     } else if (chatModelProvider && chatModel) {
       llm = chatModel.model;
+      // Set context window size for Ollama models
+      if (llm instanceof ChatOllama && body.chatModel?.provider === 'ollama') {
+        llm.numCtx = body.chatModel.ollamaContextWindow || 2048;
+      }
     }
 
     if (!llm) {

src/app/settings/page.tsx

@@ -7,6 +7,7 @@ import { Switch } from '@headlessui/react';
 import ThemeSwitcher from '@/components/theme/Switcher';
 import { ImagesIcon, VideoIcon } from 'lucide-react';
 import Link from 'next/link';
+import { PROVIDER_METADATA } from '@/lib/providers';
 
 interface SettingsType {
   chatModelProviders: {
@@ -20,10 +21,12 @@ interface SettingsType {
   anthropicApiKey: string;
   geminiApiKey: string;
   ollamaApiUrl: string;
+  lmStudioApiUrl: string;
   deepseekApiKey: string;
   customOpenaiApiKey: string;
   customOpenaiApiUrl: string;
   customOpenaiModelName: string;
+  ollamaContextWindow: number;
 }
 
 interface InputProps extends React.InputHTMLAttributes<HTMLInputElement> {
@@ -146,6 +149,11 @@ const Page = () => {
   const [automaticVideoSearch, setAutomaticVideoSearch] = useState(false);
   const [systemInstructions, setSystemInstructions] = useState<string>('');
   const [savingStates, setSavingStates] = useState<Record<string, boolean>>({});
+  const [contextWindowSize, setContextWindowSize] = useState(2048);
+  const [isCustomContextWindow, setIsCustomContextWindow] = useState(false);
+  const predefinedContextSizes = [
+    1024, 2048, 3072, 4096, 8192, 16384, 32768, 65536, 131072,
+  ];
 
   useEffect(() => {
     const fetchConfig = async () => {
@@ -157,6 +165,7 @@ const Page = () => {
       });
 
       const data = (await res.json()) as SettingsType;
+
       setConfig(data);
 
       const chatModelProvidersKeys = Object.keys(data.chatModelProviders || {});
@@ -205,6 +214,13 @@ const Page = () => {
       setAutomaticVideoSearch(
         localStorage.getItem('autoVideoSearch') === 'true',
       );
+      const storedContextWindow = parseInt(
+        localStorage.getItem('ollamaContextWindow') ?? '2048',
+      );
+      setContextWindowSize(storedContextWindow);
+      setIsCustomContextWindow(
+        !predefinedContextSizes.includes(storedContextWindow),
+      );
 
       setSystemInstructions(localStorage.getItem('systemInstructions')!);
 
@@ -364,6 +380,8 @@ const Page = () => {
         localStorage.setItem('embeddingModelProvider', value);
       } else if (key === 'embeddingModel') {
         localStorage.setItem('embeddingModel', value);
+      } else if (key === 'ollamaContextWindow') {
+        localStorage.setItem('ollamaContextWindow', value.toString());
       } else if (key === 'systemInstructions') {
         localStorage.setItem('systemInstructions', value);
       }
@@ -548,8 +566,9 @@ const Page = () => {
                         (provider) => ({
                           value: provider,
                           label:
+                            (PROVIDER_METADATA as any)[provider]?.displayName ||
                             provider.charAt(0).toUpperCase() +
-                            provider.slice(1),
+                              provider.slice(1),
                         }),
                       )}
                     />
@@ -596,6 +615,78 @@ const Page = () => {
                                 ];
                           })()}
                         />
+                        {selectedChatModelProvider === 'ollama' && (
+                          <div className="flex flex-col space-y-1">
+                            <p className="text-black/70 dark:text-white/70 text-sm">
+                              Chat Context Window Size
+                            </p>
+                            <Select
+                              value={
+                                isCustomContextWindow
+                                  ? 'custom'
+                                  : contextWindowSize.toString()
+                              }
+                              onChange={(e) => {
+                                const value = e.target.value;
+                                if (value === 'custom') {
+                                  setIsCustomContextWindow(true);
+                                } else {
+                                  setIsCustomContextWindow(false);
+                                  const numValue = parseInt(value);
+                                  setContextWindowSize(numValue);
+                                  setConfig((prev) => ({
+                                    ...prev!,
+                                    ollamaContextWindow: numValue,
+                                  }));
+                                  saveConfig('ollamaContextWindow', numValue);
+                                }
+                              }}
+                              options={[
+                                ...predefinedContextSizes.map((size) => ({
+                                  value: size.toString(),
+                                  label: `${size.toLocaleString()} tokens`,
+                                })),
+                                { value: 'custom', label: 'Custom...' },
+                              ]}
+                            />
+                            {isCustomContextWindow && (
+                              <div className="mt-2">
+                                <Input
+                                  type="number"
+                                  min={512}
+                                  value={contextWindowSize}
+                                  placeholder="Custom context window size (minimum 512)"
+                                  isSaving={savingStates['ollamaContextWindow']}
+                                  onChange={(e) => {
+                                    // Allow any value to be typed
+                                    const value =
+                                      parseInt(e.target.value) ||
+                                      contextWindowSize;
+                                    setContextWindowSize(value);
+                                  }}
+                                  onSave={(value) => {
+                                    // Validate only when saving
+                                    const numValue = Math.max(
+                                      512,
+                                      parseInt(value) || 2048,
+                                    );
+                                    setContextWindowSize(numValue);
+                                    setConfig((prev) => ({
+                                      ...prev!,
+                                      ollamaContextWindow: numValue,
+                                    }));
+                                    saveConfig('ollamaContextWindow', numValue);
+                                  }}
+                                />
+                              </div>
+                            )}
+                            <p className="text-xs text-black/60 dark:text-white/60 mt-0.5">
+                              {isCustomContextWindow
+                                ? 'Adjust the context window size for Ollama models (minimum 512 tokens)'
+                                : 'Adjust the context window size for Ollama models'}
+                            </p>
+                          </div>
+                        )}
                       </div>
                     )}
                 </div>
@@ -690,8 +781,9 @@ const Page = () => {
                         (provider) => ({
                           value: provider,
                           label:
+                            (PROVIDER_METADATA as any)[provider]?.displayName ||
                             provider.charAt(0).toUpperCase() +
-                            provider.slice(1),
+                              provider.slice(1),
                         }),
                       )}
                     />
@@ -858,6 +950,25 @@ const Page = () => {
                     onSave={(value) => saveConfig('deepseekApiKey', value)}
                   />
                 </div>
+
+                <div className="flex flex-col space-y-1">
+                  <p className="text-black/70 dark:text-white/70 text-sm">
+                    LM Studio API URL
+                  </p>
+                  <Input
+                    type="text"
+                    placeholder="LM Studio API URL"
+                    value={config.lmStudioApiUrl}
+                    isSaving={savingStates['lmStudioApiUrl']}
+                    onChange={(e) => {
+                      setConfig((prev) => ({
+                        ...prev!,
+                        lmStudioApiUrl: e.target.value,
+                      }));
+                    }}
+                    onSave={(value) => saveConfig('lmStudioApiUrl', value)}
+                  />
+                </div>
               </div>
             </SettingsSection>
           </div>

src/components/Chat.tsx

@@ -16,6 +16,8 @@ const Chat = ({
   setFileIds,
   files,
   setFiles,
+  optimizationMode,
+  setOptimizationMode,
 }: {
   messages: Message[];
   sendMessage: (message: string) => void;
@@ -26,6 +28,8 @@ const Chat = ({
   setFileIds: (fileIds: string[]) => void;
   files: File[];
   setFiles: (files: File[]) => void;
+  optimizationMode: string;
+  setOptimizationMode: (mode: string) => void;
 }) => {
   const [dividerWidth, setDividerWidth] = useState(0);
   const dividerRef = useRef<HTMLDivElement | null>(null);
@@ -99,6 +103,8 @@ const Chat = ({
             setFileIds={setFileIds}
             files={files}
             setFiles={setFiles}
+            optimizationMode={optimizationMode}
+            setOptimizationMode={setOptimizationMode}
           />
         </div>
       )}

src/components/ChatWindow.tsx

@@ -287,6 +287,16 @@ const ChatWindow = ({ id }: { id?: string }) => {
 
   const [notFound, setNotFound] = useState(false);
 
+  useEffect(() => {
+    const savedOptimizationMode = localStorage.getItem('optimizationMode');
+
+    if (savedOptimizationMode !== null) {
+      setOptimizationMode(savedOptimizationMode);
+    } else {
+      localStorage.setItem('optimizationMode', optimizationMode);
+    }
+  }, []);
+
   useEffect(() => {
     if (
       chatId &&
@@ -327,7 +337,11 @@ const ChatWindow = ({ id }: { id?: string }) => {
     }
   }, [isMessagesLoaded, isConfigReady]);
 
-  const sendMessage = async (message: string, messageId?: string) => {
+  const sendMessage = async (
+    message: string,
+    messageId?: string,
+    options?: { rewriteIndex?: number },
+  ) => {
     if (loading) return;
     if (!isConfigReady) {
       toast.error('Cannot send message before the configuration is ready');
@@ -340,6 +354,20 @@ const ChatWindow = ({ id }: { id?: string }) => {
     let sources: Document[] | undefined = undefined;
     let recievedMessage = '';
     let added = false;
+    let messageChatHistory = chatHistory;
+
+    if (options?.rewriteIndex !== undefined) {
+      const rewriteIndex = options.rewriteIndex;
+      setMessages((prev) => {
+        return [...prev.slice(0, messages.length > 2 ? rewriteIndex - 1 : 0)];
+      });
+
+      messageChatHistory = chatHistory.slice(
+        0,
+        messages.length > 2 ? rewriteIndex - 1 : 0,
+      );
+      setChatHistory(messageChatHistory);
+    }
 
     messageId = messageId ?? crypto.randomBytes(7).toString('hex');
 
@@ -363,18 +391,20 @@ const ChatWindow = ({ id }: { id?: string }) => {
 
       if (data.type === 'sources') {
         sources = data.data;
-        setMessages((prevMessages) => [
-          ...prevMessages,
-          {
-            content: '',
-            messageId: data.messageId,
-            chatId: chatId!,
-            role: 'assistant',
-            sources: sources,
-            createdAt: new Date(),
-          },
-        ]);
-        added = true;
+        if (!added) {
+          setMessages((prevMessages) => [
+            ...prevMessages,
+            {
+              content: '',
+              messageId: data.messageId,
+              chatId: chatId!,
+              role: 'assistant',
+              sources: sources,
+              createdAt: new Date(),
+            },
+          ]);
+          added = true;
+        }
         setMessageAppeared(true);
       }
 
@@ -392,20 +422,20 @@ const ChatWindow = ({ id }: { id?: string }) => {
             },
           ]);
           added = true;
-          setMessageAppeared(true);
-        } else {
-          setMessages((prev) =>
-            prev.map((message) => {
-              if (message.messageId === data.messageId) {
-                return { ...message, content: message.content + data.data };
-              }
-
-              return message;
-            }),
-          );
         }
 
+        setMessages((prev) =>
+          prev.map((message) => {
+            if (message.messageId === data.messageId) {
+              return { ...message, content: message.content + data.data };
+            }
+
+            return message;
+          }),
+        );
+
         recievedMessage += data.data;
+        setMessageAppeared(true);
       }
 
       if (data.type === 'messageEnd') {
@@ -453,6 +483,9 @@ const ChatWindow = ({ id }: { id?: string }) => {
       }
     };
 
+    const ollamaContextWindow =
+      localStorage.getItem('ollamaContextWindow') || '2048';
+
     const res = await fetch('/api/chat', {
       method: 'POST',
       headers: {
@@ -469,10 +502,13 @@ const ChatWindow = ({ id }: { id?: string }) => {
         files: fileIds,
         focusMode: focusMode,
         optimizationMode: optimizationMode,
-        history: chatHistory,
+        history: messageChatHistory,
         chatModel: {
           name: chatModelProvider.name,
           provider: chatModelProvider.provider,
+          ...(chatModelProvider.provider === 'ollama' && {
+            ollamaContextWindow: parseInt(ollamaContextWindow),
+          }),
         },
         embeddingModel: {
           name: embeddingModelProvider.name,
@@ -510,20 +546,13 @@ const ChatWindow = ({ id }: { id?: string }) => {
   };
 
   const rewrite = (messageId: string) => {
-    const index = messages.findIndex((msg) => msg.messageId === messageId);
-
-    if (index === -1) return;
-
-    const message = messages[index - 1];
-
-    setMessages((prev) => {
-      return [...prev.slice(0, messages.length > 2 ? index - 1 : 0)];
+    const messageIndex = messages.findIndex(
+      (msg) => msg.messageId === messageId,
+    );
+    if (messageIndex == -1) return;
+    sendMessage(messages[messageIndex - 1].content, messageId, {
+      rewriteIndex: messageIndex,
     });
-    setChatHistory((prev) => {
-      return [...prev.slice(0, messages.length > 2 ? index - 1 : 0)];
-    });
-
-    sendMessage(message.content, message.messageId);
   };
 
   useEffect(() => {
@@ -568,6 +597,8 @@ const ChatWindow = ({ id }: { id?: string }) => {
               setFileIds={setFileIds}
               files={files}
               setFiles={setFiles}
+              optimizationMode={optimizationMode}
+              setOptimizationMode={setOptimizationMode}
             />
           </>
         ) : (

src/components/MessageBox.tsx

@@ -97,6 +97,7 @@ const MessageBox = ({
           },
         ),
       );
+      setSpeechMessage(message.content.replace(regex, ''));
       return;
     }
 

src/components/MessageInput.tsx

@@ -4,6 +4,7 @@ import { useEffect, useRef, useState } from 'react';
 import TextareaAutosize from 'react-textarea-autosize';
 import Attach from './MessageInputActions/Attach';
 import CopilotToggle from './MessageInputActions/Copilot';
+import Optimization from './MessageInputActions/Optimization';
 import { File } from './ChatWindow';
 import AttachSmall from './MessageInputActions/AttachSmall';
 
@@ -14,6 +15,8 @@ const MessageInput = ({
   setFileIds,
   files,
   setFiles,
+  optimizationMode,
+  setOptimizationMode,
 }: {
   sendMessage: (message: string) => void;
   loading: boolean;
@@ -21,6 +24,8 @@ const MessageInput = ({
   setFileIds: (fileIds: string[]) => void;
   files: File[];
   setFiles: (files: File[]) => void;
+  optimizationMode: string;
+  setOptimizationMode: (mode: string) => void;
 }) => {
   const [copilotEnabled, setCopilotEnabled] = useState(false);
   const [message, setMessage] = useState('');
@@ -40,20 +45,16 @@ const MessageInput = ({
   useEffect(() => {
     const handleKeyDown = (e: KeyboardEvent) => {
       const activeElement = document.activeElement;
-
       const isInputFocused =
         activeElement?.tagName === 'INPUT' ||
         activeElement?.tagName === 'TEXTAREA' ||
         activeElement?.hasAttribute('contenteditable');
-
       if (e.key === '/' && !isInputFocused) {
         e.preventDefault();
         inputRef.current?.focus();
       }
     };
-
     document.addEventListener('keydown', handleKeyDown);
-
     return () => {
       document.removeEventListener('keydown', handleKeyDown);
     };
@@ -75,58 +76,95 @@ const MessageInput = ({
         }
       }}
       className={cn(
-        'bg-light-secondary dark:bg-dark-secondary p-4 flex items-center overflow-hidden border border-light-200 dark:border-dark-200',
-        mode === 'multi' ? 'flex-col rounded-lg' : 'flex-row rounded-full',
+        'bg-light-secondary dark:bg-dark-secondary p-4 flex items-center border border-light-200 dark:border-dark-200',
+        mode === 'multi'
+          ? 'flex-col rounded-lg'
+          : 'flex-col md:flex-row rounded-lg md:rounded-full',
       )}
     >
       {mode === 'single' && (
-        <AttachSmall
-          fileIds={fileIds}
-          setFileIds={setFileIds}
-          files={files}
-          setFiles={setFiles}
-        />
-      )}
-      <TextareaAutosize
-        ref={inputRef}
-        value={message}
-        onChange={(e) => setMessage(e.target.value)}
-        onHeightChange={(height, props) => {
-          setTextareaRows(Math.ceil(height / props.rowHeight));
-        }}
-        className="transition bg-transparent dark:placeholder:text-white/50 placeholder:text-sm text-sm dark:text-white resize-none focus:outline-none w-full px-2 max-h-24 lg:max-h-36 xl:max-h-48 flex-grow flex-shrink"
-        placeholder="Ask a follow-up"
-      />
-      {mode === 'single' && (
-        <div className="flex flex-row items-center space-x-4">
-          <CopilotToggle
-            copilotEnabled={copilotEnabled}
-            setCopilotEnabled={setCopilotEnabled}
-          />
-          <button
-            disabled={message.trim().length === 0 || loading}
-            className="bg-[#24A0ED] text-white disabled:text-black/50 dark:disabled:text-white/50 hover:bg-opacity-85 transition duration-100 disabled:bg-[#e0e0dc79] dark:disabled:bg-[#ececec21] rounded-full p-2"
-          >
-            <ArrowUp className="bg-background" size={17} />
-          </button>
-        </div>
-      )}
-      {mode === 'multi' && (
-        <div className="flex flex-row items-center justify-between w-full pt-2">
-          <AttachSmall
-            fileIds={fileIds}
-            setFileIds={setFileIds}
-            files={files}
-            setFiles={setFiles}
-          />
-          <div className="flex flex-row items-center space-x-4">
+        <div className="flex flex-row items-center justify-between w-full mb-2 md:mb-0 md:w-auto">
+          <div className="flex flex-row items-center space-x-2">
+            <AttachSmall
+              fileIds={fileIds}
+              setFileIds={setFileIds}
+              files={files}
+              setFiles={setFiles}
+            />
+            <Optimization
+              optimizationMode={optimizationMode}
+              setOptimizationMode={setOptimizationMode}
+            />
+          </div>
+          <div className="md:hidden">
             <CopilotToggle
               copilotEnabled={copilotEnabled}
               setCopilotEnabled={setCopilotEnabled}
             />
+          </div>
+        </div>
+      )}
+      <div className="flex flex-row items-center w-full">
+        <TextareaAutosize
+          ref={inputRef}
+          value={message}
+          onChange={(e) => setMessage(e.target.value)}
+          onHeightChange={(height, props) => {
+            setTextareaRows(Math.ceil(height / props.rowHeight));
+          }}
+          className="transition bg-transparent dark:placeholder:text-white/50 placeholder:text-sm text-sm dark:text-white resize-none focus:outline-none w-full px-2 max-h-24 lg:max-h-36 xl:max-h-48 flex-grow flex-shrink"
+          placeholder="Ask a follow-up"
+        />
+        {mode === 'single' && (
+          <div className="flex flex-row items-center space-x-4">
+            <div className="hidden md:block">
+              <CopilotToggle
+                copilotEnabled={copilotEnabled}
+                setCopilotEnabled={setCopilotEnabled}
+              />
+            </div>
             <button
               disabled={message.trim().length === 0 || loading}
-              className="bg-[#24A0ED] text-white text-black/50 dark:disabled:text-white/50 hover:bg-opacity-85 transition duration-100 disabled:bg-[#e0e0dc79] dark:disabled:bg-[#ececec21] rounded-full p-2"
+              className="bg-[#24A0ED] text-white disabled:text-black/50 dark:disabled:text-white/50 hover:bg-opacity-85 transition duration-100 disabled:bg-[#e0e0dc79] dark:disabled:bg-[#ececec21] rounded-full p-2"
+            >
+              <ArrowUp className="bg-background" size={17} />
+            </button>
+          </div>
+        )}
+      </div>
+
+      {mode === 'multi' && (
+        <div className="flex flex-col md:flex-row items-start md:items-center justify-between w-full pt-2">
+          <div className="flex flex-row items-center justify-between w-full md:w-auto mb-2 md:mb-0">
+            <div className="flex flex-row items-center space-x-2">
+              <AttachSmall
+                fileIds={fileIds}
+                setFileIds={setFileIds}
+                files={files}
+                setFiles={setFiles}
+              />
+              <Optimization
+                optimizationMode={optimizationMode}
+                setOptimizationMode={setOptimizationMode}
+              />
+            </div>
+            <div className="md:hidden">
+              <CopilotToggle
+                copilotEnabled={copilotEnabled}
+                setCopilotEnabled={setCopilotEnabled}
+              />
+            </div>
+          </div>
+          <div className="flex flex-row items-center space-x-4 self-end">
+            <div className="hidden md:block">
+              <CopilotToggle
+                copilotEnabled={copilotEnabled}
+                setCopilotEnabled={setCopilotEnabled}
+              />
+            </div>
+            <button
+              disabled={message.trim().length === 0 || loading}
+              className="bg-[#24A0ED] text-white disabled:text-black/50 dark:disabled:text-white/50 hover:bg-opacity-85 transition duration-100 disabled:bg-[#e0e0dc79] dark:disabled:bg-[#ececec21] rounded-full p-2"
             >
               <ArrowUp className="bg-background" size={17} />
             </button>

src/components/MessageInputActions/Optimization.tsx

@@ -1,4 +1,4 @@
-import { ChevronDown, Sliders, Star, Zap } from 'lucide-react';
+import { ChevronDown, Minimize2, Sliders, Star, Zap } from 'lucide-react';
 import { cn } from '@/lib/utils';
 import {
   Popover,
@@ -7,7 +7,6 @@ import {
   Transition,
 } from '@headlessui/react';
 import { Fragment } from 'react';
-
 const OptimizationModes = [
   {
     key: 'speed',
@@ -41,8 +40,13 @@ const Optimization = ({
   optimizationMode: string;
   setOptimizationMode: (mode: string) => void;
 }) => {
+  const handleOptimizationChange = (mode: string) => {
+    setOptimizationMode(mode);
+    localStorage.setItem('optimizationMode', mode);
+  };
+
   return (
-    <Popover className="relative w-full max-w-[15rem] md:max-w-md lg:max-w-lg">
+    <Popover className="relative">
       <PopoverButton
         type="button"
         className="p-2 text-black/50 dark:text-white/50 rounded-xl hover:bg-light-secondary dark:hover:bg-dark-secondary active:scale-95 transition duration-200 hover:text-black dark:hover:text-white"
@@ -70,17 +74,19 @@ const Optimization = ({
         leaveFrom="opacity-100 translate-y-0"
         leaveTo="opacity-0 translate-y-1"
       >
-        <PopoverPanel className="absolute z-10 w-64 md:w-[250px] right-0">
-          <div className="flex flex-col gap-2 bg-light-primary dark:bg-dark-primary border rounded-lg border-light-200 dark:border-dark-200 w-full p-4 max-h-[200px] md:max-h-none overflow-y-auto">
+        <PopoverPanel className="absolute z-10 bottom-[100%] mb-2 left-1/2 transform -translate-x-1/2">
+          <div className="flex flex-col gap-2 bg-light-primary dark:bg-dark-primary border rounded-lg border-light-200 dark:border-dark-200 w-max max-w-[300px] p-4 max-h-[200px] md:max-h-none overflow-y-auto">
             {OptimizationModes.map((mode, i) => (
               <PopoverButton
-                onClick={() => setOptimizationMode(mode.key)}
+                onClick={() => handleOptimizationChange(mode.key)}
                 key={i}
+                disabled={mode.key === 'quality'}
                 className={cn(
                   'p-2 rounded-lg flex flex-col items-start justify-start text-start space-y-1 duration-200 cursor-pointer transition',
                   optimizationMode === mode.key
                     ? 'bg-light-secondary dark:bg-dark-secondary'
                     : 'hover:bg-light-secondary dark:hover:bg-dark-secondary',
+                  mode.key === 'quality' && 'opacity-50 cursor-not-allowed',
                 )}
               >
                 <div className="flex flex-row items-center space-x-1 text-black dark:text-white">

src/components/SearchImages.tsx

@@ -35,9 +35,10 @@ const SearchImages = ({
 
             const chatModelProvider = localStorage.getItem('chatModelProvider');
             const chatModel = localStorage.getItem('chatModel');
-
             const customOpenAIBaseURL = localStorage.getItem('openAIBaseURL');
             const customOpenAIKey = localStorage.getItem('openAIApiKey');
+            const ollamaContextWindow =
+              localStorage.getItem('ollamaContextWindow') || '2048';
 
             const res = await fetch(`/api/images`, {
               method: 'POST',
@@ -54,6 +55,9 @@ const SearchImages = ({
                     customOpenAIBaseURL: customOpenAIBaseURL,
                     customOpenAIKey: customOpenAIKey,
                   }),
+                  ...(chatModelProvider === 'ollama' && {
+                    ollamaContextWindow: parseInt(ollamaContextWindow),
+                  }),
                 },
               }),
             });

src/components/SearchVideos.tsx

@@ -50,9 +50,10 @@ const Searchvideos = ({
 
             const chatModelProvider = localStorage.getItem('chatModelProvider');
             const chatModel = localStorage.getItem('chatModel');
-
             const customOpenAIBaseURL = localStorage.getItem('openAIBaseURL');
             const customOpenAIKey = localStorage.getItem('openAIApiKey');
+            const ollamaContextWindow =
+              localStorage.getItem('ollamaContextWindow') || '2048';
 
             const res = await fetch(`/api/videos`, {
               method: 'POST',
@@ -69,6 +70,9 @@ const Searchvideos = ({
                     customOpenAIBaseURL: customOpenAIBaseURL,
                     customOpenAIKey: customOpenAIKey,
                   }),
+                  ...(chatModelProvider === 'ollama' && {
+                    ollamaContextWindow: parseInt(ollamaContextWindow),
+                  }),
                 },
               }),
             });

src/lib/actions.ts

@@ -6,6 +6,8 @@ export const getSuggestions = async (chatHisory: Message[]) => {
 
   const customOpenAIKey = localStorage.getItem('openAIApiKey');
   const customOpenAIBaseURL = localStorage.getItem('openAIBaseURL');
+  const ollamaContextWindow =
+    localStorage.getItem('ollamaContextWindow') || '2048';
 
   const res = await fetch(`/api/suggestions`, {
     method: 'POST',
@@ -21,6 +23,9 @@ export const getSuggestions = async (chatHisory: Message[]) => {
           customOpenAIKey,
           customOpenAIBaseURL,
         }),
+        ...(chatModelProvider === 'ollama' && {
+          ollamaContextWindow: parseInt(ollamaContextWindow),
+        }),
       },
     }),
   });

src/lib/config.ts

@@ -1,7 +1,14 @@
-import fs from 'fs';
-import path from 'path';
 import toml from '@iarna/toml';
 
+// Use dynamic imports for Node.js modules to prevent client-side errors
+let fs: any;
+let path: any;
+if (typeof window === 'undefined') {
+  // We're on the server
+  fs = require('fs');
+  path = require('path');
+}
+
 const configFileName = 'config.toml';
 
 interface Config {
@@ -28,6 +35,9 @@ interface Config {
     DEEPSEEK: {
       API_KEY: string;
     };
+    LM_STUDIO: {
+      API_URL: string;
+    };
     CUSTOM_OPENAI: {
       API_URL: string;
       API_KEY: string;
@@ -43,10 +53,17 @@ type RecursivePartial<T> = {
   [P in keyof T]?: RecursivePartial<T[P]>;
 };
 
-const loadConfig = () =>
-  toml.parse(
-    fs.readFileSync(path.join(process.cwd(), `${configFileName}`), 'utf-8'),
-  ) as any as Config;
+const loadConfig = () => {
+  // Server-side only
+  if (typeof window === 'undefined') {
+    return toml.parse(
+      fs.readFileSync(path.join(process.cwd(), `${configFileName}`), 'utf-8'),
+    ) as any as Config;
+  }
+
+  // Client-side fallback - settings will be loaded via API
+  return {} as Config;
+};
 
 export const getSimilarityMeasure = () =>
   loadConfig().GENERAL.SIMILARITY_MEASURE;
@@ -77,6 +94,9 @@ export const getCustomOpenaiApiUrl = () =>
 export const getCustomOpenaiModelName = () =>
   loadConfig().MODELS.CUSTOM_OPENAI.MODEL_NAME;
 
+export const getLMStudioApiEndpoint = () =>
+  loadConfig().MODELS.LM_STUDIO.API_URL;
+
 const mergeConfigs = (current: any, update: any): any => {
   if (update === null || update === undefined) {
     return current;
@@ -109,10 +129,13 @@ const mergeConfigs = (current: any, update: any): any => {
 };
 
 export const updateConfig = (config: RecursivePartial<Config>) => {
-  const currentConfig = loadConfig();
-  const mergedConfig = mergeConfigs(currentConfig, config);
-  fs.writeFileSync(
-    path.join(path.join(process.cwd(), `${configFileName}`)),
-    toml.stringify(mergedConfig),
-  );
+  // Server-side only
+  if (typeof window === 'undefined') {
+    const currentConfig = loadConfig();
+    const mergedConfig = mergeConfigs(currentConfig, config);
+    fs.writeFileSync(
+      path.join(path.join(process.cwd(), `${configFileName}`)),
+      toml.stringify(mergedConfig),
+    );
+  }
 };

src/lib/providers/anthropic.ts

@@ -1,6 +1,11 @@
 import { ChatAnthropic } from '@langchain/anthropic';
 import { ChatModel } from '.';
 import { getAnthropicApiKey } from '../config';
+
+export const PROVIDER_INFO = {
+  key: 'anthropic',
+  displayName: 'Anthropic',
+};
 import { BaseChatModel } from '@langchain/core/language_models/chat_models';
 
 const anthropicChatModels: Record<string, string>[] = [

src/lib/providers/deepseek.ts

@@ -3,6 +3,11 @@ import { getDeepseekApiKey } from '../config';
 import { ChatModel } from '.';
 import { BaseChatModel } from '@langchain/core/language_models/chat_models';
 
+export const PROVIDER_INFO = {
+  key: 'deepseek',
+  displayName: 'Deepseek AI',
+};
+
 const deepseekChatModels: Record<string, string>[] = [
   {
     displayName: 'Deepseek Chat (Deepseek V3)',

src/lib/providers/gemini.ts

@@ -4,6 +4,11 @@ import {
 } from '@langchain/google-genai';
 import { getGeminiApiKey } from '../config';
 import { ChatModel, EmbeddingModel } from '.';
+
+export const PROVIDER_INFO = {
+  key: 'gemini',
+  displayName: 'Google Gemini',
+};
 import { BaseChatModel } from '@langchain/core/language_models/chat_models';
 import { Embeddings } from '@langchain/core/embeddings';
 

src/lib/providers/groq.ts

@@ -1,6 +1,11 @@
 import { ChatOpenAI } from '@langchain/openai';
 import { getGroqApiKey } from '../config';
 import { ChatModel } from '.';
+
+export const PROVIDER_INFO = {
+  key: 'groq',
+  displayName: 'Groq',
+};
 import { BaseChatModel } from '@langchain/core/language_models/chat_models';
 
 const groqChatModels: Record<string, string>[] = [

src/lib/providers/index.ts

@@ -1,18 +1,60 @@
 import { Embeddings } from '@langchain/core/embeddings';
 import { BaseChatModel } from '@langchain/core/language_models/chat_models';
-import { loadOpenAIChatModels, loadOpenAIEmbeddingModels } from './openai';
+import {
+  loadOpenAIChatModels,
+  loadOpenAIEmbeddingModels,
+  PROVIDER_INFO as OpenAIInfo,
+  PROVIDER_INFO,
+} from './openai';
 import {
   getCustomOpenaiApiKey,
   getCustomOpenaiApiUrl,
   getCustomOpenaiModelName,
 } from '../config';
 import { ChatOpenAI } from '@langchain/openai';
-import { loadOllamaChatModels, loadOllamaEmbeddingModels } from './ollama';
-import { loadGroqChatModels } from './groq';
-import { loadAnthropicChatModels } from './anthropic';
-import { loadGeminiChatModels, loadGeminiEmbeddingModels } from './gemini';
-import { loadTransformersEmbeddingsModels } from './transformers';
-import { loadDeepseekChatModels } from './deepseek';
+import {
+  loadOllamaChatModels,
+  loadOllamaEmbeddingModels,
+  PROVIDER_INFO as OllamaInfo,
+} from './ollama';
+import { loadGroqChatModels, PROVIDER_INFO as GroqInfo } from './groq';
+import {
+  loadAnthropicChatModels,
+  PROVIDER_INFO as AnthropicInfo,
+} from './anthropic';
+import {
+  loadGeminiChatModels,
+  loadGeminiEmbeddingModels,
+  PROVIDER_INFO as GeminiInfo,
+} from './gemini';
+import {
+  loadTransformersEmbeddingsModels,
+  PROVIDER_INFO as TransformersInfo,
+} from './transformers';
+import {
+  loadDeepseekChatModels,
+  PROVIDER_INFO as DeepseekInfo,
+} from './deepseek';
+import {
+  loadLMStudioChatModels,
+  loadLMStudioEmbeddingsModels,
+  PROVIDER_INFO as LMStudioInfo,
+} from './lmstudio';
+
+export const PROVIDER_METADATA = {
+  openai: OpenAIInfo,
+  ollama: OllamaInfo,
+  groq: GroqInfo,
+  anthropic: AnthropicInfo,
+  gemini: GeminiInfo,
+  transformers: TransformersInfo,
+  deepseek: DeepseekInfo,
+  lmstudio: LMStudioInfo,
+  custom_openai: {
+    key: 'custom_openai',
+    displayName: 'Custom OpenAI',
+  },
+};
 
 export interface ChatModel {
   displayName: string;
@@ -34,6 +76,7 @@ export const chatModelProviders: Record<
   anthropic: loadAnthropicChatModels,
   gemini: loadGeminiChatModels,
   deepseek: loadDeepseekChatModels,
+  lmstudio: loadLMStudioChatModels,
 };
 
 export const embeddingModelProviders: Record<
@@ -44,6 +87,7 @@ export const embeddingModelProviders: Record<
   ollama: loadOllamaEmbeddingModels,
   gemini: loadGeminiEmbeddingModels,
   transformers: loadTransformersEmbeddingsModels,
+  lmstudio: loadLMStudioEmbeddingsModels,
 };
 
 export const getAvailableChatModelProviders = async () => {

src/lib/providers/lmstudio.ts

@@ -0,0 +1,100 @@
+import { getKeepAlive, getLMStudioApiEndpoint } from '../config';
+import axios from 'axios';
+import { ChatModel, EmbeddingModel } from '.';
+
+export const PROVIDER_INFO = {
+  key: 'lmstudio',
+  displayName: 'LM Studio',
+};
+import { ChatOpenAI } from '@langchain/openai';
+import { OpenAIEmbeddings } from '@langchain/openai';
+import { BaseChatModel } from '@langchain/core/language_models/chat_models';
+import { Embeddings } from '@langchain/core/embeddings';
+
+interface LMStudioModel {
+  id: string;
+  name?: string;
+}
+
+const ensureV1Endpoint = (endpoint: string): string =>
+  endpoint.endsWith('/v1') ? endpoint : `${endpoint}/v1`;
+
+const checkServerAvailability = async (endpoint: string): Promise<boolean> => {
+  try {
+    await axios.get(`${ensureV1Endpoint(endpoint)}/models`, {
+      headers: { 'Content-Type': 'application/json' },
+    });
+    return true;
+  } catch {
+    return false;
+  }
+};
+
+export const loadLMStudioChatModels = async () => {
+  const endpoint = getLMStudioApiEndpoint();
+
+  if (!endpoint) return {};
+  if (!(await checkServerAvailability(endpoint))) return {};
+
+  try {
+    const response = await axios.get(`${ensureV1Endpoint(endpoint)}/models`, {
+      headers: { 'Content-Type': 'application/json' },
+    });
+
+    const chatModels: Record<string, ChatModel> = {};
+
+    response.data.data.forEach((model: LMStudioModel) => {
+      chatModels[model.id] = {
+        displayName: model.name || model.id,
+        model: new ChatOpenAI({
+          openAIApiKey: 'lm-studio',
+          configuration: {
+            baseURL: ensureV1Endpoint(endpoint),
+          },
+          modelName: model.id,
+          temperature: 0.7,
+          streaming: true,
+          maxRetries: 3,
+        }) as unknown as BaseChatModel,
+      };
+    });
+
+    return chatModels;
+  } catch (err) {
+    console.error(`Error loading LM Studio models: ${err}`);
+    return {};
+  }
+};
+
+export const loadLMStudioEmbeddingsModels = async () => {
+  const endpoint = getLMStudioApiEndpoint();
+
+  if (!endpoint) return {};
+  if (!(await checkServerAvailability(endpoint))) return {};
+
+  try {
+    const response = await axios.get(`${ensureV1Endpoint(endpoint)}/models`, {
+      headers: { 'Content-Type': 'application/json' },
+    });
+
+    const embeddingsModels: Record<string, EmbeddingModel> = {};
+
+    response.data.data.forEach((model: LMStudioModel) => {
+      embeddingsModels[model.id] = {
+        displayName: model.name || model.id,
+        model: new OpenAIEmbeddings({
+          openAIApiKey: 'lm-studio',
+          configuration: {
+            baseURL: ensureV1Endpoint(endpoint),
+          },
+          modelName: model.id,
+        }) as unknown as Embeddings,
+      };
+    });
+
+    return embeddingsModels;
+  } catch (err) {
+    console.error(`Error loading LM Studio embeddings model: ${err}`);
+    return {};
+  }
+};

src/lib/providers/ollama.ts

@@ -1,8 +1,13 @@
 import axios from 'axios';
 import { getKeepAlive, getOllamaApiEndpoint } from '../config';
 import { ChatModel, EmbeddingModel } from '.';
-import { ChatOllama } from '@langchain/community/chat_models/ollama';
-import { OllamaEmbeddings } from '@langchain/community/embeddings/ollama';
+
+export const PROVIDER_INFO = {
+  key: 'ollama',
+  displayName: 'Ollama',
+};
+import { ChatOllama } from '@langchain/ollama';
+import { OllamaEmbeddings } from '@langchain/ollama';
 
 export const loadOllamaChatModels = async () => {
   const ollamaApiEndpoint = getOllamaApiEndpoint();

src/lib/providers/openai.ts

@@ -1,6 +1,11 @@
 import { ChatOpenAI, OpenAIEmbeddings } from '@langchain/openai';
 import { getOpenaiApiKey } from '../config';
 import { ChatModel, EmbeddingModel } from '.';
+
+export const PROVIDER_INFO = {
+  key: 'openai',
+  displayName: 'OpenAI',
+};
 import { BaseChatModel } from '@langchain/core/language_models/chat_models';
 import { Embeddings } from '@langchain/core/embeddings';
 

src/lib/providers/transformers.ts

@@ -1,5 +1,10 @@
 import { HuggingFaceTransformersEmbeddings } from '../huggingfaceTransformer';
 
+export const PROVIDER_INFO = {
+  key: 'transformers',
+  displayName: 'Hugging Face',
+};
+
 export const loadTransformersEmbeddingsModels = async () => {
   try {
     const embeddingModels = {

src/lib/search/metaSearchAgent.ts

@@ -6,20 +6,24 @@ import {
   MessagesPlaceholder,
   PromptTemplate,
 } from '@langchain/core/prompts';
+import {
+  RunnableLambda,
+  RunnableMap,
+  RunnableSequence,
+} from '@langchain/core/runnables';
 import { BaseMessage } from '@langchain/core/messages';
 import { StringOutputParser } from '@langchain/core/output_parsers';
 import LineListOutputParser from '../outputParsers/listLineOutputParser';
 import LineOutputParser from '../outputParsers/lineOutputParser';
 import { getDocumentsFromLinks } from '../utils/documents';
 import { Document } from 'langchain/document';
-import { searchSearxng, SearxngSearchResult } from '../searxng';
+import { searchSearxng } from '../searxng';
 import path from 'node:path';
 import fs from 'node:fs';
 import computeSimilarity from '../utils/computeSimilarity';
 import formatChatHistoryAsString from '../utils/formatHistory';
 import eventEmitter from 'events';
 import { StreamEvent } from '@langchain/core/tracers/log_stream';
-import { EventEmitter } from 'node:stream';
 
 export interface MetaSearchAgentType {
   searchAndAnswer: (
@@ -43,7 +47,7 @@ interface Config {
   activeEngines: string[];
 }
 
-type SearchInput = {
+type BasicChainInput = {
   chat_history: BaseMessage[];
   query: string;
 };
@@ -56,385 +60,237 @@ class MetaSearchAgent implements MetaSearchAgentType {
     this.config = config;
   }
 
-  private async searchSources(
-    llm: BaseChatModel,
-    input: SearchInput,
-    emitter: EventEmitter,
-  ) {
+  private async createSearchRetrieverChain(llm: BaseChatModel) {
     (llm as unknown as ChatOpenAI).temperature = 0;
 
-    const chatPrompt = PromptTemplate.fromTemplate(
-      this.config.queryGeneratorPrompt,
-    );
+    return RunnableSequence.from([
+      PromptTemplate.fromTemplate(this.config.queryGeneratorPrompt),
+      llm,
+      this.strParser,
+      RunnableLambda.from(async (input: string) => {
+        const linksOutputParser = new LineListOutputParser({
+          key: 'links',
+        });
 
-    const processedChatPrompt = await chatPrompt.invoke({
-      chat_history: formatChatHistoryAsString(input.chat_history),
-      query: input.query,
-    });
+        const questionOutputParser = new LineOutputParser({
+          key: 'question',
+        });
 
-    const llmRes = await llm.invoke(processedChatPrompt);
-    const messageStr = await this.strParser.invoke(llmRes);
+        const links = await linksOutputParser.parse(input);
+        let question = this.config.summarizer
+          ? await questionOutputParser.parse(input)
+          : input;
 
-    const linksOutputParser = new LineListOutputParser({
-      key: 'links',
-    });
-
-    const questionOutputParser = new LineOutputParser({
-      key: 'question',
-    });
-
-    const links = await linksOutputParser.parse(messageStr);
-    let question = this.config.summarizer
-      ? await questionOutputParser.parse(messageStr)
-      : messageStr;
-
-    if (question === 'not_needed') {
-      return { query: '', docs: [] };
-    }
-
-    if (links.length > 0) {
-      if (question.length === 0) {
-        question = 'summarize';
-      }
-
-      let docs: Document[] = [];
-
-      const linkDocs = await getDocumentsFromLinks({ links });
-
-      const docGroups: Document[] = [];
-
-      linkDocs.map((doc) => {
-        const URLDocExists = docGroups.find(
-          (d) =>
-            d.metadata.url === doc.metadata.url && d.metadata.totalDocs < 10,
-        );
-
-        if (!URLDocExists) {
-          docGroups.push({
-            ...doc,
-            metadata: {
-              ...doc.metadata,
-              totalDocs: 1,
-            },
-          });
+        if (question === 'not_needed') {
+          return { query: '', docs: [] };
         }
 
-        const docIndex = docGroups.findIndex(
-          (d) =>
-            d.metadata.url === doc.metadata.url && d.metadata.totalDocs < 10,
-        );
+        if (links.length > 0) {
+          if (question.length === 0) {
+            question = 'summarize';
+          }
 
-        if (docIndex !== -1) {
-          docGroups[docIndex].pageContent =
-            docGroups[docIndex].pageContent + `\n\n` + doc.pageContent;
-          docGroups[docIndex].metadata.totalDocs += 1;
-        }
-      });
+          let docs: Document[] = [];
 
-      await Promise.all(
-        docGroups.map(async (doc) => {
-          const res = await llm.invoke(`
-        You are a web search summarizer, tasked with summarizing a piece of text retrieved from a web search. Your job is to summarize the 
-        text into a detailed, 2-4 paragraph explanation that captures the main ideas and provides a comprehensive answer to the query.
-        If the query is \"summarize\", you should provide a detailed summary of the text. If the query is a specific question, you should answer it in the summary.
-        
-        - **Journalistic tone**: The summary should sound professional and journalistic, not too casual or vague.
-        - **Thorough and detailed**: Ensure that every key point from the text is captured and that the summary directly answers the query.
-        - **Not too lengthy, but detailed**: The summary should be informative but not excessively long. Focus on providing detailed information in a concise format.
+          const linkDocs = await getDocumentsFromLinks({ links });
 
-        The text will be shared inside the \`text\` XML tag, and the query inside the \`query\` XML tag.
+          const docGroups: Document[] = [];
 
-        <example>
-        1. \`<text>
-        Docker is a set of platform-as-a-service products that use OS-level virtualization to deliver software in packages called containers. 
-        It was first released in 2013 and is developed by Docker, Inc. Docker is designed to make it easier to create, deploy, and run applications 
-        by using containers.
-        </text>
+          linkDocs.map((doc) => {
+            const URLDocExists = docGroups.find(
+              (d) =>
+                d.metadata.url === doc.metadata.url &&
+                d.metadata.totalDocs < 10,
+            );
 
-        <query>
-        What is Docker and how does it work?
-        </query>
+            if (!URLDocExists) {
+              docGroups.push({
+                ...doc,
+                metadata: {
+                  ...doc.metadata,
+                  totalDocs: 1,
+                },
+              });
+            }
 
-        Response:
-        Docker is a revolutionary platform-as-a-service product developed by Docker, Inc., that uses container technology to make application 
-        deployment more efficient. It allows developers to package their software with all necessary dependencies, making it easier to run in 
-        any environment. Released in 2013, Docker has transformed the way applications are built, deployed, and managed.
-        \`
-        2. \`<text>
-        The theory of relativity, or simply relativity, encompasses two interrelated theories of Albert Einstein: special relativity and general
-        relativity. However, the word "relativity" is sometimes used in reference to Galilean invariance. The term "theory of relativity" was based
-        on the expression "relative theory" used by Max Planck in 1906. The theory of relativity usually encompasses two interrelated theories by
-        Albert Einstein: special relativity and general relativity. Special relativity applies to all physical phenomena in the absence of gravity.
-        General relativity explains the law of gravitation and its relation to other forces of nature. It applies to the cosmological and astrophysical
-        realm, including astronomy.
-        </text>
+            const docIndex = docGroups.findIndex(
+              (d) =>
+                d.metadata.url === doc.metadata.url &&
+                d.metadata.totalDocs < 10,
+            );
 
-        <query>
-        summarize
-        </query>
-
-        Response:
-        The theory of relativity, developed by Albert Einstein, encompasses two main theories: special relativity and general relativity. Special
-        relativity applies to all physical phenomena in the absence of gravity, while general relativity explains the law of gravitation and its
-        relation to other forces of nature. The theory of relativity is based on the concept of "relative theory," as introduced by Max Planck in
-        1906. It is a fundamental theory in physics that has revolutionized our understanding of the universe.
-        \`
-        </example>
-
-        Everything below is the actual data you will be working with. Good luck!
-
-        <query>
-        ${question}
-        </query>
-
-        <text>
-        ${doc.pageContent}
-        </text>
-
-        Make sure to answer the query in the summary.
-      `);
-
-          const document = new Document({
-            pageContent: res.content as string,
-            metadata: {
-              title: doc.metadata.title,
-              url: doc.metadata.url,
-            },
+            if (docIndex !== -1) {
+              docGroups[docIndex].pageContent =
+                docGroups[docIndex].pageContent + `\n\n` + doc.pageContent;
+              docGroups[docIndex].metadata.totalDocs += 1;
+            }
           });
 
-          docs.push(document);
-        }),
-      );
+          await Promise.all(
+            docGroups.map(async (doc) => {
+              const res = await llm.invoke(`
+            You are a web search summarizer, tasked with summarizing a piece of text retrieved from a web search. Your job is to summarize the 
+            text into a detailed, 2-4 paragraph explanation that captures the main ideas and provides a comprehensive answer to the query.
+            If the query is \"summarize\", you should provide a detailed summary of the text. If the query is a specific question, you should answer it in the summary.
+            
+            - **Journalistic tone**: The summary should sound professional and journalistic, not too casual or vague.
+            - **Thorough and detailed**: Ensure that every key point from the text is captured and that the summary directly answers the query.
+            - **Not too lengthy, but detailed**: The summary should be informative but not excessively long. Focus on providing detailed information in a concise format.
 
-      return { query: question, docs: docs };
-    } else {
-      question = question.replace(/<think>.*?<\/think>/g, '');
+            The text will be shared inside the \`text\` XML tag, and the query inside the \`query\` XML tag.
 
-      const res = await searchSearxng(question, {
-        language: 'en',
-        engines: this.config.activeEngines,
-      });
+            <example>
+            1. \`<text>
+            Docker is a set of platform-as-a-service products that use OS-level virtualization to deliver software in packages called containers. 
+            It was first released in 2013 and is developed by Docker, Inc. Docker is designed to make it easier to create, deploy, and run applications 
+            by using containers.
+            </text>
 
-      const documents = res.results.map(
-        (result) =>
-          new Document({
-            pageContent:
-              result.content ||
-              (this.config.activeEngines.includes('youtube')
-                ? result.title
-                : '') /* Todo: Implement transcript grabbing using Youtubei (source: https://www.npmjs.com/package/youtubei) */,
-            metadata: {
-              title: result.title,
-              url: result.url,
-              ...(result.img_src && { img_src: result.img_src }),
-            },
-          }),
-      );
+            <query>
+            What is Docker and how does it work?
+            </query>
 
-      return { query: question, docs: documents };
-    }
+            Response:
+            Docker is a revolutionary platform-as-a-service product developed by Docker, Inc., that uses container technology to make application 
+            deployment more efficient. It allows developers to package their software with all necessary dependencies, making it easier to run in 
+            any environment. Released in 2013, Docker has transformed the way applications are built, deployed, and managed.
+            \`
+            2. \`<text>
+            The theory of relativity, or simply relativity, encompasses two interrelated theories of Albert Einstein: special relativity and general
+            relativity. However, the word "relativity" is sometimes used in reference to Galilean invariance. The term "theory of relativity" was based
+            on the expression "relative theory" used by Max Planck in 1906. The theory of relativity usually encompasses two interrelated theories by
+            Albert Einstein: special relativity and general relativity. Special relativity applies to all physical phenomena in the absence of gravity.
+            General relativity explains the law of gravitation and its relation to other forces of nature. It applies to the cosmological and astrophysical
+            realm, including astronomy.
+            </text>
+
+            <query>
+            summarize
+            </query>
+
+            Response:
+            The theory of relativity, developed by Albert Einstein, encompasses two main theories: special relativity and general relativity. Special
+            relativity applies to all physical phenomena in the absence of gravity, while general relativity explains the law of gravitation and its
+            relation to other forces of nature. The theory of relativity is based on the concept of "relative theory," as introduced by Max Planck in
+            1906. It is a fundamental theory in physics that has revolutionized our understanding of the universe.
+            \`
+            </example>
+
+            Everything below is the actual data you will be working with. Good luck!
+
+            <query>
+            ${question}
+            </query>
+
+            <text>
+            ${doc.pageContent}
+            </text>
+
+            Make sure to answer the query in the summary.
+          `);
+
+              const document = new Document({
+                pageContent: res.content as string,
+                metadata: {
+                  title: doc.metadata.title,
+                  url: doc.metadata.url,
+                },
+              });
+
+              docs.push(document);
+            }),
+          );
+
+          return { query: question, docs: docs };
+        } else {
+          question = question.replace(/<think>.*?<\/think>/g, '');
+
+          const res = await searchSearxng(question, {
+            language: 'en',
+            engines: this.config.activeEngines,
+          });
+
+          const documents = res.results.map(
+            (result) =>
+              new Document({
+                pageContent:
+                  result.content ||
+                  (this.config.activeEngines.includes('youtube')
+                    ? result.title
+                    : '') /* Todo: Implement transcript grabbing using Youtubei (source: https://www.npmjs.com/package/youtubei) */,
+                metadata: {
+                  title: result.title,
+                  url: result.url,
+                  ...(result.img_src && { img_src: result.img_src }),
+                },
+              }),
+          );
+
+          return { query: question, docs: documents };
+        }
+      }),
+    ]);
   }
 
-  private async performDeepResearch(
-    llm: BaseChatModel,
-    input: SearchInput,
-    emitter: EventEmitter,
-  ) {
-    (llm as unknown as ChatOpenAI).temperature = 0;
-
-    const queryGenPrompt = PromptTemplate.fromTemplate(
-      this.config.queryGeneratorPrompt,
-    );
-
-    const formattedChatPrompt = await queryGenPrompt.invoke({
-      chat_history: formatChatHistoryAsString(input.chat_history),
-      query: input.query,
-    });
-
-    let i = 0;
-    let currentQuery = await this.strParser.invoke(
-      await llm.invoke(formattedChatPrompt),
-    );
-    const originalQuery = currentQuery;
-    const pastQueries: string[] = [];
-    const results: SearxngSearchResult[] = [];
-
-    while (i < 10) {
-      const res = await searchSearxng(currentQuery, {
-        language: 'en',
-        engines: this.config.activeEngines,
-      });
-
-      results.push(...res.results);
-
-      const reflectorPrompt = PromptTemplate.fromTemplate(`
-          You are an LLM that is tasked with reflecting on the results of a search query.
-
-          ## Goal
-          You will be given question of the user, a list of search results collected from the web to answer that question along with past queries made to collect those results. You have to analyze the results based on user's question and do the following:
-          
-          1. Identify unexplored areas or areas with less detailed information in the results and generate a new query that focuses on those areas. The new queries should be more specific and a similar query shall not exist in past queries which will be provided to you. Make sure to include keywords that you're looking for because the new query will be used to search the web for information on that topic. Make sure the query contains only 1 question and is not too long to ensure it is Search Engine friendly.
-          2. You'll have to generate a description explaining what you are doing for example "I am looking for more information about X" or "Understanding how X works" etc. The description should be short and concise.
-          
-          ## Output format
-
-          You need to output in XML format and do not generate any other text. ake sure to not include any other text in the output or start a conversation in the output. The output should be in the following format:
-
-          <query>(query)</query>
-          <description>(description)</description>
-
-          ## Example
-          Say the user asked "What is Llama 4 by Meta?" and let search results contain information about Llama 4 being an LLM and very little information about its features. You can output:
-
-          <query>Llama 4 features</query> // Generate queries that capture keywords for SEO and not making words like "How", "What", "Why" etc.
-          <description>Looking for new features in Llama 4</description>
-
-          or something like
-
-          <query>How is Llama 4 better than its previous generation models</query>
-          <description>Understanding the difference between Llama 4 and previous generation models.</description>
-          
-          ## BELOW IS THE ACTUAL DATA YOU WILL BE WORKING WITH. IT IS NOT A PART OF EXAMPLES. YOU'LL HAVE TO GENERATE YOUR ANSWER BASED ON THIS DATA.
-          <user_question>\n{question}\n</user_question>
-          <search_results>\n{search_results}\n</search_results>
-          <past_queries>\n{past_queries}\n</past_queries>
-
-          Response:
-      `);
-
-      const formattedReflectorPrompt = await reflectorPrompt.invoke({
-        question: originalQuery,
-        search_results: results
-          .map(
-            (result) => `<result>${result.title} - ${result.content}</result>`,
-          )
-          .join('\n'),
-        past_queries: pastQueries.map((q) => `<query>${q}</query>`).join('\n'),
-      });
-
-      const feedback = await this.strParser.invoke(
-        await llm.invoke(formattedReflectorPrompt),
-      );
-
-      console.log(`Feedback: ${feedback}`);
-
-      const queryOutputParser = new LineOutputParser({
-        key: 'query',
-      });
-
-      const descriptionOutputParser = new LineOutputParser({
-        key: 'description',
-      });
-
-      currentQuery = await queryOutputParser.parse(feedback);
-      const description = await descriptionOutputParser.parse(feedback);
-      console.log(`Query: ${currentQuery}`);
-      console.log(`Description: ${description}`);
-
-      pastQueries.push(currentQuery);
-      ++i;
-    }
-
-    const uniqueResults: SearxngSearchResult[] = [];
-
-    results.forEach((res) => {
-      const exists = uniqueResults.find((r) => r.url === res.url);
-
-      if (!exists) {
-        uniqueResults.push(res);
-      } else {
-        exists.content += `\n\n` + res.content;
-      }
-    });
-
-    const documents = uniqueResults /* .slice(0, 50) */
-      .map(
-        (r) =>
-          new Document({
-            pageContent: r.content || '',
-            metadata: {
-              title: r.title,
-              url: r.url,
-              ...(r.img_src && { img_src: r.img_src }),
-            },
-          }),
-      );
-
-    return documents;
-  }
-
-  private async streamAnswer(
+  private async createAnsweringChain(
     llm: BaseChatModel,
     fileIds: string[],
     embeddings: Embeddings,
     optimizationMode: 'speed' | 'balanced' | 'quality',
     systemInstructions: string,
-    input: SearchInput,
-    emitter: EventEmitter,
   ) {
-    const chatPrompt = ChatPromptTemplate.fromMessages([
-      ['system', this.config.responsePrompt],
-      new MessagesPlaceholder('chat_history'),
-      ['user', '{query}'],
-    ]);
+    return RunnableSequence.from([
+      RunnableMap.from({
+        systemInstructions: () => systemInstructions,
+        query: (input: BasicChainInput) => input.query,
+        chat_history: (input: BasicChainInput) => input.chat_history,
+        date: () => new Date().toISOString(),
+        context: RunnableLambda.from(async (input: BasicChainInput) => {
+          const processedHistory = formatChatHistoryAsString(
+            input.chat_history,
+          );
 
-    let context = '';
+          let docs: Document[] | null = null;
+          let query = input.query;
 
-    if (optimizationMode === 'speed' || optimizationMode === 'balanced') {
-      let docs: Document[] | null = null;
-      let query = input.query;
+          if (this.config.searchWeb) {
+            const searchRetrieverChain =
+              await this.createSearchRetrieverChain(llm);
 
-      if (this.config.searchWeb) {
-        const searchResults = await this.searchSources(llm, input, emitter);
+            const searchRetrieverResult = await searchRetrieverChain.invoke({
+              chat_history: processedHistory,
+              query,
+            });
 
-        query = searchResults.query;
-        docs = searchResults.docs;
-      }
+            query = searchRetrieverResult.query;
+            docs = searchRetrieverResult.docs;
+          }
 
-      const sortedDocs = await this.rerankDocs(
-        query,
-        docs ?? [],
-        fileIds,
-        embeddings,
-        optimizationMode,
-      );
+          const sortedDocs = await this.rerankDocs(
+            query,
+            docs ?? [],
+            fileIds,
+            embeddings,
+            optimizationMode,
+          );
 
-      emitter.emit(
-        'data',
-        JSON.stringify({ type: 'sources', data: sortedDocs }),
-      );
-
-      context = this.processDocs(sortedDocs);
-    } else if (optimizationMode === 'quality') {
-      let docs: Document[] = [];
-
-      docs = await this.performDeepResearch(llm, input, emitter);
-
-      emitter.emit('data', JSON.stringify({ type: 'sources', data: docs }));
-
-      context = this.processDocs(docs);
-    }
-
-    const formattedChatPrompt = await chatPrompt.invoke({
-      query: input.query,
-      chat_history: input.chat_history,
-      date: new Date().toISOString(),
-      context: context,
-      systemInstructions: systemInstructions,
+          return sortedDocs;
+        })
+          .withConfig({
+            runName: 'FinalSourceRetriever',
+          })
+          .pipe(this.processDocs),
+      }),
+      ChatPromptTemplate.fromMessages([
+        ['system', this.config.responsePrompt],
+        new MessagesPlaceholder('chat_history'),
+        ['user', '{query}'],
+      ]),
+      llm,
+      this.strParser,
+    ]).withConfig({
+      runName: 'FinalResponseGenerator',
     });
-
-    const llmRes = await llm.stream(formattedChatPrompt);
-
-    for await (const data of llmRes) {
-      const messageStr = await this.strParser.invoke(data);
-
-      emitter.emit(
-        'data',
-        JSON.stringify({ type: 'response', data: messageStr }),
-      );
-    }
-
-    emitter.emit('end');
   }
 
   private async rerankDocs(
@@ -570,13 +426,44 @@ class MetaSearchAgent implements MetaSearchAgentType {
     return docs
       .map(
         (_, index) =>
-          `${index + 1}. ${docs[index].metadata.title} ${
-            docs[index].pageContent
-          }`,
+          `${index + 1}. ${docs[index].metadata.title} ${docs[index].pageContent}`,
       )
       .join('\n');
   }
 
+  private async handleStream(
+    stream: AsyncGenerator<StreamEvent, any, any>,
+    emitter: eventEmitter,
+  ) {
+    for await (const event of stream) {
+      if (
+        event.event === 'on_chain_end' &&
+        event.name === 'FinalSourceRetriever'
+      ) {
+        ``;
+        emitter.emit(
+          'data',
+          JSON.stringify({ type: 'sources', data: event.data.output }),
+        );
+      }
+      if (
+        event.event === 'on_chain_stream' &&
+        event.name === 'FinalResponseGenerator'
+      ) {
+        emitter.emit(
+          'data',
+          JSON.stringify({ type: 'response', data: event.data.chunk }),
+        );
+      }
+      if (
+        event.event === 'on_chain_end' &&
+        event.name === 'FinalResponseGenerator'
+      ) {
+        emitter.emit('end');
+      }
+    }
+  }
+
   async searchAndAnswer(
     message: string,
     history: BaseMessage[],
@@ -588,19 +475,26 @@ class MetaSearchAgent implements MetaSearchAgentType {
   ) {
     const emitter = new eventEmitter();
 
-    this.streamAnswer(
+    const answeringChain = await this.createAnsweringChain(
       llm,
       fileIds,
       embeddings,
       optimizationMode,
       systemInstructions,
+    );
+
+    const stream = answeringChain.streamEvents(
       {
         chat_history: history,
         query: message,
       },
-      emitter,
+      {
+        version: 'v1',
+      },
     );
 
+    this.handleStream(stream, emitter);
+
     return emitter;
   }
 }

src/lib/searxng.ts

@@ -8,7 +8,7 @@ interface SearxngSearchOptions {
   pageno?: number;
 }
 
-export interface SearxngSearchResult {
+interface SearxngSearchResult {
   title: string;
   url: string;
   img_src?: string;