Merge 590a52d38c into a8d410bc2f

Larger instances will benefit from this change massively.
Also QWant was spamming the logs with some chaptcha problem so best to disable it for now.

docker-compose.yaml

@@ -26,9 +26,21 @@ services:
       - ./config.toml:/home/perplexica/config.toml
     restart: unless-stopped
 
+  redict:
+    image: registry.redict.io/redict:latest
+    container_name: perplexica-redict
+    ports:
+      - "6379:6379"
+    volumes:
+      - redict_data:/data
+    networks:
+      - perplexica-network
+    restart: unless-stopped
+
 networks:
   perplexica-network:
 
 volumes:
   backend-dbstore:
   uploads:
+  redict_data:

docs/API/SEARCH.md

@@ -33,7 +33,6 @@ The API accepts a JSON object in the request body, where you define the focus mo
     ["human", "Hi, how are you?"],
     ["assistant", "I am doing well, how can I help you today?"]
   ],
-  "systemInstructions": "Focus on providing technical details about Perplexica's architecture.",
   "stream": false
 }
 ```
@@ -64,8 +63,6 @@ The API accepts a JSON object in the request body, where you define the focus mo
 
 - **`query`** (string, required): The search query or question.
 
-- **`systemInstructions`** (string, optional): Custom instructions provided by the user to guide the AI's response. These instructions are treated as user preferences and have lower priority than the system's core instructions. For example, you can specify a particular writing style, format, or focus area.
-
 - **`history`** (array, optional): An array of message pairs representing the conversation history. Each pair consists of a role (either 'human' or 'assistant') and the message content. This allows the system to use the context of the conversation to refine results. Example:
 
   ```json

package.json

@@ -1,6 +1,6 @@
 {
   "name": "perplexica-frontend",
-  "version": "1.10.2",
+  "version": "1.10.1",
   "license": "MIT",
   "author": "ItzCrazyKns",
   "scripts": {

sample.config.toml

@@ -22,8 +22,5 @@ MODEL_NAME = ""
 [MODELS.OLLAMA]
 API_URL = "" # Ollama API URL - http://host.docker.internal:11434
 
-[MODELS.DEEPSEEK]
-API_KEY = ""
-
 [API_ENDPOINTS]
 SEARXNG = "" # SearxNG API URL - http://localhost:32768

searxng/settings.yml

@@ -12,6 +12,11 @@ search:
 server:
   secret_key: 'a2fb23f1b02e6ee83875b09826990de0f6bd908b6638e8c10277d415f6ab852b' # Is overwritten by ${SEARXNG_SECRET}
 
+redis:
+  url: redis://redict:6379/0
+
 engines:
   - name: wolframalpha
     disabled: false
+  - name: qwant
+    disabled: true

src/app/api/config/route.ts

@@ -7,7 +7,6 @@ import {
   getGroqApiKey,
   getOllamaApiEndpoint,
   getOpenaiApiKey,
-  getDeepseekApiKey,
   updateConfig,
 } from '@/lib/config';
 import {
@@ -54,7 +53,6 @@ export const GET = async (req: Request) => {
     config['anthropicApiKey'] = getAnthropicApiKey();
     config['groqApiKey'] = getGroqApiKey();
     config['geminiApiKey'] = getGeminiApiKey();
-    config['deepseekApiKey'] = getDeepseekApiKey();
     config['customOpenaiApiUrl'] = getCustomOpenaiApiUrl();
     config['customOpenaiApiKey'] = getCustomOpenaiApiKey();
     config['customOpenaiModelName'] = getCustomOpenaiModelName();
@@ -90,9 +88,6 @@ export const POST = async (req: Request) => {
         OLLAMA: {
           API_URL: config.ollamaApiUrl,
         },
-        DEEPSEEK: {
-          API_KEY: config.deepseekApiKey,
-        },
         CUSTOM_OPENAI: {
           API_URL: config.customOpenaiApiUrl,
           API_KEY: config.customOpenaiApiKey,

src/app/api/search/route.ts

@@ -34,7 +34,6 @@ interface ChatRequestBody {
   query: string;
   history: Array<[string, string]>;
   stream?: boolean;
-  systemInstructions?: string;
 }
 
 export const POST = async (req: Request) => {
@@ -126,7 +125,7 @@ export const POST = async (req: Request) => {
       embeddings,
       body.optimizationMode,
       [],
-      body.systemInstructions || '',
+      "",
     );
 
     if (!body.stream) {

src/app/settings/page.tsx

@@ -20,7 +20,6 @@ interface SettingsType {
   anthropicApiKey: string;
   geminiApiKey: string;
   ollamaApiUrl: string;
-  deepseekApiKey: string;
   customOpenaiApiKey: string;
   customOpenaiApiUrl: string;
   customOpenaiModelName: string;
@@ -839,25 +838,6 @@ const Page = () => {
                     onSave={(value) => saveConfig('geminiApiKey', value)}
                   />
                 </div>
-
-                <div className="flex flex-col space-y-1">
-                  <p className="text-black/70 dark:text-white/70 text-sm">
-                    Deepseek API Key
-                  </p>
-                  <Input
-                    type="text"
-                    placeholder="Deepseek API Key"
-                    value={config.deepseekApiKey}
-                    isSaving={savingStates['deepseekApiKey']}
-                    onChange={(e) => {
-                      setConfig((prev) => ({
-                        ...prev!,
-                        deepseekApiKey: e.target.value,
-                      }));
-                    }}
-                    onSave={(value) => saveConfig('deepseekApiKey', value)}
-                  />
-                </div>
               </div>
             </SettingsSection>
           </div>

src/components/ChatWindow.tsx

@@ -363,18 +363,20 @@ 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 +394,20 @@ 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') {

src/components/MessageBox.tsx

@@ -48,7 +48,6 @@ const MessageBox = ({
   const [speechMessage, setSpeechMessage] = useState(message.content);
 
   useEffect(() => {
-    const citationRegex = /\[([^\]]+)\]/g;
     const regex = /\[(\d+)\]/g;
     let processedMessage = message.content;
 
@@ -68,33 +67,11 @@ const MessageBox = ({
     ) {
       setParsedMessage(
         processedMessage.replace(
-          citationRegex,
-          (_, capturedContent: string) => {
-            const numbers = capturedContent
-              .split(',')
-              .map((numStr) => numStr.trim());
-
-            const linksHtml = numbers
-              .map((numStr) => {
-                const number = parseInt(numStr);
-
-                if (isNaN(number) || number <= 0) {
-                  return `[${numStr}]`;
-                }
-
-                const source = message.sources?.[number - 1];
-                const url = source?.metadata?.url;
-
-                if (url) {
-                  return `<a href="${url}" target="_blank" className="bg-light-secondary dark:bg-dark-secondary px-1 rounded ml-1 no-underline text-xs text-black/70 dark:text-white/70 relative">${numStr}</a>`;
-                } else {
-                  return `[${numStr}]`;
-                }
-              })
-              .join('');
-
-            return linksHtml;
-          },
+          regex,
+          (_, number) =>
+            `<a href="${
+              message.sources?.[number - 1]?.metadata?.url
+            }" target="_blank" className="bg-light-secondary dark:bg-dark-secondary px-1 rounded ml-1 no-underline text-xs text-black/70 dark:text-white/70 relative">${number}</a>`,
         ),
       );
       return;

src/components/MessageInputActions/Optimization.tsx

@@ -76,11 +76,13 @@ const Optimization = ({
               <PopoverButton
                 onClick={() => setOptimizationMode(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/lib/config.ts

@@ -25,9 +25,6 @@ interface Config {
     OLLAMA: {
       API_URL: string;
     };
-    DEEPSEEK: {
-      API_KEY: string;
-    };
     CUSTOM_OPENAI: {
       API_URL: string;
       API_KEY: string;
@@ -66,8 +63,6 @@ export const getSearxngApiEndpoint = () =>
 
 export const getOllamaApiEndpoint = () => loadConfig().MODELS.OLLAMA.API_URL;
 
-export const getDeepseekApiKey = () => loadConfig().MODELS.DEEPSEEK.API_KEY;
-
 export const getCustomOpenaiApiKey = () =>
   loadConfig().MODELS.CUSTOM_OPENAI.API_KEY;
 

src/lib/prompts/webSearch.ts

@@ -1,6 +1,6 @@
 export const webSearchRetrieverPrompt = `
 You are an AI question rephraser. You will be given a conversation and a follow-up question,  you will have to rephrase the follow up question so it is a standalone question and can be used by another LLM to search the web for information to answer it.
-If it is a simple writing task or a greeting (unless the greeting contains a question after it) like Hi, Hello, How are you, etc. than a question then you need to return \`not_needed\` as the response (This is because the LLM won't need to search the web for finding information on this topic).
+If it is a smple writing task or a greeting (unless the greeting contains a question after it) like Hi, Hello, How are you, etc. than a question then you need to return \`not_needed\` as the response (This is because the LLM won't need to search the web for finding information on this topic).
 If the user asks some question from some URL or wants you to summarize a PDF or a webpage (via URL) you need to return the links inside the \`links\` XML block and the question inside the \`question\` XML block. If the user wants to you to summarize the webpage or the PDF you need to return \`summarize\` inside the \`question\` XML block in place of a question and the link to summarize in the \`links\` XML block.
 You must always return the rephrased question inside the \`question\` XML block, if there are no links in the follow-up question then don't insert a \`links\` XML block in your response.
 

src/lib/providers/deepseek.ts

@@ -1,44 +0,0 @@
-import { ChatOpenAI } from '@langchain/openai';
-import { getDeepseekApiKey } from '../config';
-import { ChatModel } from '.';
-import { BaseChatModel } from '@langchain/core/language_models/chat_models';
-
-const deepseekChatModels: Record<string, string>[] = [
-  {
-    displayName: 'Deepseek Chat (Deepseek V3)',
-    key: 'deepseek-chat',
-  },
-  {
-    displayName: 'Deepseek Reasoner (Deepseek R1)',
-    key: 'deepseek-reasoner',
-  },
-];
-
-export const loadDeepseekChatModels = async () => {
-  const deepseekApiKey = getDeepseekApiKey();
-
-  if (!deepseekApiKey) return {};
-
-  try {
-    const chatModels: Record<string, ChatModel> = {};
-
-    deepseekChatModels.forEach((model) => {
-      chatModels[model.key] = {
-        displayName: model.displayName,
-        model: new ChatOpenAI({
-          openAIApiKey: deepseekApiKey,
-          modelName: model.key,
-          temperature: 0.7,
-          configuration: {
-            baseURL: 'https://api.deepseek.com',
-          },
-        }) as unknown as BaseChatModel,
-      };
-    });
-
-    return chatModels;
-  } catch (err) {
-    console.error(`Error loading Deepseek models: ${err}`);
-    return {};
-  }
-};

src/lib/providers/gemini.ts

@@ -40,12 +40,8 @@ const geminiChatModels: Record<string, string>[] = [
 
 const geminiEmbeddingModels: Record<string, string>[] = [
   {
-    displayName: 'Text Embedding 004',
-    key: 'models/text-embedding-004',
-  },
-  {
-    displayName: 'Embedding 001',
-    key: 'models/embedding-001',
+    displayName: 'Gemini Embedding',
+    key: 'gemini-embedding-exp',
   },
 ];
 

src/lib/providers/groq.ts

@@ -72,14 +72,6 @@ const groqChatModels: Record<string, string>[] = [
     displayName: 'Llama 3.2 90B Vision Preview (Preview)',
     key: 'llama-3.2-90b-vision-preview',
   },
-  /* {
-    displayName: 'Llama 4 Maverick 17B 128E Instruct (Preview)',
-    key: 'meta-llama/llama-4-maverick-17b-128e-instruct',
-  }, */
-  {
-    displayName: 'Llama 4 Scout 17B 16E Instruct (Preview)',
-    key: 'meta-llama/llama-4-scout-17b-16e-instruct',
-  },
 ];
 
 export const loadGroqChatModels = async () => {

src/lib/providers/index.ts

@@ -12,7 +12,6 @@ import { loadGroqChatModels } from './groq';
 import { loadAnthropicChatModels } from './anthropic';
 import { loadGeminiChatModels, loadGeminiEmbeddingModels } from './gemini';
 import { loadTransformersEmbeddingsModels } from './transformers';
-import { loadDeepseekChatModels } from './deepseek';
 
 export interface ChatModel {
   displayName: string;
@@ -33,7 +32,6 @@ export const chatModelProviders: Record<
   groq: loadGroqChatModels,
   anthropic: loadAnthropicChatModels,
   gemini: loadGeminiChatModels,
-  deepseek: loadDeepseekChatModels,
 };
 
 export const embeddingModelProviders: Record<

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;