Merge pull request #736 from ItzCrazyKns/master

Merge master into feat/deep-research

README.md

@@ -89,7 +89,6 @@ There are mainly 2 ways of installing Perplexica - With Docker, Without Docker.
    - `OPENAI`: Your OpenAI API key. **You only need to fill this if you wish to use OpenAI's models**.
    - `OLLAMA`: Your Ollama API URL. You should enter it as `http://host.docker.internal:PORT_NUMBER`. If you installed Ollama on port 11434, use `http://host.docker.internal:11434`. For other ports, adjust accordingly. **You need to fill this if you wish to use Ollama's models instead of OpenAI's**.
    - `GROQ`: Your Groq API key. **You only need to fill this if you wish to use Groq's hosted models**.
-   - `OPENROUTER`: Your OpenRouter API key. **You only need to fill this if you wish to use models via OpenRouter**.
    - `ANTHROPIC`: Your Anthropic API key. **You only need to fill this if you wish to use Anthropic models**.
 
      **Note**: You can change these after starting Perplexica from the settings dialog.

docs/API/SEARCH.md

@@ -33,6 +33,7 @@ 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
 }
 ```
@@ -63,6 +64,8 @@ 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": "1.10.1",
+  "version": "1.10.2",
   "license": "MIT",
   "author": "ItzCrazyKns",
   "scripts": {

sample.config.toml

@@ -11,9 +11,6 @@ API_KEY = ""
 [MODELS.ANTHROPIC]
 API_KEY = ""
 
-[MODELS.OPENROUTER]
-API_KEY = ""
-
 [MODELS.GEMINI]
 API_KEY = ""
 
@@ -25,5 +22,8 @@ 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

src/app/api/config/route.ts

@@ -5,9 +5,9 @@ import {
   getCustomOpenaiModelName,
   getGeminiApiKey,
   getGroqApiKey,
-  getOpenrouterApiKey,
   getOllamaApiEndpoint,
   getOpenaiApiKey,
+  getDeepseekApiKey,
   updateConfig,
 } from '@/lib/config';
 import {
@@ -53,8 +53,8 @@ export const GET = async (req: Request) => {
     config['ollamaApiUrl'] = getOllamaApiEndpoint();
     config['anthropicApiKey'] = getAnthropicApiKey();
     config['groqApiKey'] = getGroqApiKey();
-    config['openrouterApiKey'] = getOpenrouterApiKey();
     config['geminiApiKey'] = getGeminiApiKey();
+    config['deepseekApiKey'] = getDeepseekApiKey();
     config['customOpenaiApiUrl'] = getCustomOpenaiApiUrl();
     config['customOpenaiApiKey'] = getCustomOpenaiApiKey();
     config['customOpenaiModelName'] = getCustomOpenaiModelName();
@@ -81,9 +81,6 @@ export const POST = async (req: Request) => {
         GROQ: {
           API_KEY: config.groqApiKey,
         },
-        OPENROUTER: {
-          API_KEY: config.openrouterApiKey,
-        },
         ANTHROPIC: {
           API_KEY: config.anthropicApiKey,
         },
@@ -93,6 +90,9 @@ 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

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

src/app/settings/page.tsx

@@ -17,10 +17,10 @@ interface SettingsType {
   };
   openaiApiKey: string;
   groqApiKey: string;
-  openrouterApiKey: string;
   anthropicApiKey: string;
   geminiApiKey: string;
   ollamaApiUrl: string;
+  deepseekApiKey: string;
   customOpenaiApiKey: string;
   customOpenaiApiUrl: string;
   customOpenaiModelName: string;
@@ -802,25 +802,6 @@ const Page = () => {
                   />
                 </div>
 
-                <div className="flex flex-col space-y-1">
-                  <p className="text-black/70 dark:text-white/70 text-sm">
-                    OpenRouter API Key
-                  </p>
-                  <Input
-                    type="text"
-                    placeholder="OpenRouter API Key"
-                    value={config.openrouterApiKey}
-                    isSaving={savingStates['openrouterApiKey']}
-                    onChange={(e) => {
-                      setConfig((prev) => ({
-                        ...prev!,
-                        openrouterApiKey: e.target.value,
-                      }));
-                    }}
-                    onSave={(value) => saveConfig('openrouterApiKey', value)}
-                  />
-                </div>
-
                 <div className="flex flex-col space-y-1">
                   <p className="text-black/70 dark:text-white/70 text-sm">
                     Anthropic API Key
@@ -858,6 +839,25 @@ 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

@@ -363,7 +363,6 @@ const ChatWindow = ({ id }: { id?: string }) => {
 
       if (data.type === 'sources') {
         sources = data.data;
-        if (!added) {
         setMessages((prevMessages) => [
           ...prevMessages,
           {
@@ -376,7 +375,6 @@ const ChatWindow = ({ id }: { id?: string }) => {
           },
         ]);
         added = true;
-        }
         setMessageAppeared(true);
       }
 
@@ -394,8 +392,8 @@ const ChatWindow = ({ id }: { id?: string }) => {
             },
           ]);
           added = true;
-        }
+          setMessageAppeared(true);
-
+        } else {
           setMessages((prev) =>
             prev.map((message) => {
               if (message.messageId === data.messageId) {
@@ -405,9 +403,9 @@ const ChatWindow = ({ id }: { id?: string }) => {
               return message;
             }),
           );
+        }
 
         recievedMessage += data.data;
-        setMessageAppeared(true);
       }
 
       if (data.type === 'messageEnd') {

src/components/MessageBox.tsx

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

src/components/MessageInputActions/Optimization.tsx

@@ -76,13 +76,11 @@ 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

@@ -25,7 +25,7 @@ interface Config {
     OLLAMA: {
       API_URL: string;
     };
-    OPENROUTER: {
+    DEEPSEEK: {
       API_KEY: string;
     };
     CUSTOM_OPENAI: {
@@ -57,8 +57,6 @@ export const getOpenaiApiKey = () => loadConfig().MODELS.OPENAI.API_KEY;
 
 export const getGroqApiKey = () => loadConfig().MODELS.GROQ.API_KEY;
 
-export const getOpenrouterApiKey = () => loadConfig().MODELS.OPENROUTER.API_KEY;
-
 export const getAnthropicApiKey = () => loadConfig().MODELS.ANTHROPIC.API_KEY;
 
 export const getGeminiApiKey = () => loadConfig().MODELS.GEMINI.API_KEY;
@@ -68,6 +66,8 @@ 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 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 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 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

@@ -0,0 +1,44 @@
+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

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

src/lib/providers/groq.ts

@@ -72,6 +72,14 @@ 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 +12,7 @@ import { loadGroqChatModels } from './groq';
 import { loadAnthropicChatModels } from './anthropic';
 import { loadGeminiChatModels, loadGeminiEmbeddingModels } from './gemini';
 import { loadTransformersEmbeddingsModels } from './transformers';
-import { loadOpenrouterChatModels } from '@/lib/providers/openrouter';
+import { loadDeepseekChatModels } from './deepseek';
 
 export interface ChatModel {
   displayName: string;
@@ -33,7 +33,7 @@ export const chatModelProviders: Record<
   groq: loadGroqChatModels,
   anthropic: loadAnthropicChatModels,
   gemini: loadGeminiChatModels,
-  openrouter: loadOpenrouterChatModels,
+  deepseek: loadDeepseekChatModels,
 };
 
 export const embeddingModelProviders: Record<

src/lib/providers/openrouter.ts

@@ -1,61 +0,0 @@
-import { ChatOpenAI } from '@langchain/openai';
-import { getOpenrouterApiKey } from '../config';
-import { ChatModel } from '.';
-import { BaseChatModel } from '@langchain/core/language_models/chat_models';
-
-let openrouterChatModels: Record<string, string>[] = [];
-
-async function fetchModelList(): Promise<void> {
-  try {
-    const response = await fetch('https://openrouter.ai/api/v1/models', {
-      method: 'GET',
-      headers: {
-        'Content-Type': 'application/json',
-      },
-    });
-
-    if (!response.ok) {
-      throw new Error(`API request failed with status: ${response.status}`);
-    }
-
-    const data = await response.json();
-
-    openrouterChatModels = data.data.map((model: any) => ({
-      displayName: model.name,
-      key: model.id,
-    }));
-  } catch (error) {
-    console.error('Error fetching models:', error);
-  }
-}
-
-export const loadOpenrouterChatModels = async () => {
-  await fetchModelList();
-
-  const openrouterApikey = getOpenrouterApiKey();
-
-  if (!openrouterApikey) return {};
-
-  try {
-    const chatModels: Record<string, ChatModel> = {};
-
-    openrouterChatModels.forEach((model) => {
-      chatModels[model.key] = {
-        displayName: model.displayName,
-        model: new ChatOpenAI({
-          openAIApiKey: openrouterApikey,
-          modelName: model.key,
-          temperature: 0.7,
-          configuration: {
-            baseURL: 'https://openrouter.ai/api/v1',
-          },
-        }) as unknown as BaseChatModel,
-      };
-    });
-
-    return chatModels;
-  } catch (err) {
-    console.error(`Error loading Openrouter models: ${err}`);
-    return {};
-  }
-};

src/lib/search/metaSearchAgent.ts

@@ -6,24 +6,20 @@ 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 } from '../searxng';
+import { searchSearxng, SearxngSearchResult } 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: (
@@ -47,7 +43,7 @@ interface Config {
   activeEngines: string[];
 }
 
-type BasicChainInput = {
+type SearchInput = {
   chat_history: BaseMessage[];
   query: string;
 };
@@ -60,14 +56,25 @@ class MetaSearchAgent implements MetaSearchAgentType {
     this.config = config;
   }
 
-  private async createSearchRetrieverChain(llm: BaseChatModel) {
+  private async searchSources(
+    llm: BaseChatModel,
+    input: SearchInput,
+    emitter: EventEmitter,
+  ) {
     (llm as unknown as ChatOpenAI).temperature = 0;
 
-    return RunnableSequence.from([
+    const chatPrompt = PromptTemplate.fromTemplate(
-      PromptTemplate.fromTemplate(this.config.queryGeneratorPrompt),
+      this.config.queryGeneratorPrompt,
-      llm,
+    );
-      this.strParser,
+
-      RunnableLambda.from(async (input: string) => {
+    const processedChatPrompt = await chatPrompt.invoke({
+      chat_history: formatChatHistoryAsString(input.chat_history),
+      query: input.query,
+    });
+
+    const llmRes = await llm.invoke(processedChatPrompt);
+    const messageStr = await this.strParser.invoke(llmRes);
+
     const linksOutputParser = new LineListOutputParser({
       key: 'links',
     });
@@ -76,10 +83,10 @@ class MetaSearchAgent implements MetaSearchAgentType {
       key: 'question',
     });
 
-        const links = await linksOutputParser.parse(input);
+    const links = await linksOutputParser.parse(messageStr);
     let question = this.config.summarizer
-          ? await questionOutputParser.parse(input)
+      ? await questionOutputParser.parse(messageStr)
-          : input;
+      : messageStr;
 
     if (question === 'not_needed') {
       return { query: '', docs: [] };
@@ -99,8 +106,7 @@ class MetaSearchAgent implements MetaSearchAgentType {
       linkDocs.map((doc) => {
         const URLDocExists = docGroups.find(
           (d) =>
-                d.metadata.url === doc.metadata.url &&
+            d.metadata.url === doc.metadata.url && d.metadata.totalDocs < 10,
-                d.metadata.totalDocs < 10,
         );
 
         if (!URLDocExists) {
@@ -115,8 +121,7 @@ class MetaSearchAgent implements MetaSearchAgentType {
 
         const docIndex = docGroups.findIndex(
           (d) =>
-                d.metadata.url === doc.metadata.url &&
+            d.metadata.url === doc.metadata.url && d.metadata.totalDocs < 10,
-                d.metadata.totalDocs < 10,
         );
 
         if (docIndex !== -1) {
@@ -228,42 +233,162 @@ class MetaSearchAgent implements MetaSearchAgentType {
 
       return { query: question, docs: documents };
     }
-      }),
-    ]);
   }
 
-  private async createAnsweringChain(
+  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(
     llm: BaseChatModel,
     fileIds: string[],
     embeddings: Embeddings,
     optimizationMode: 'speed' | 'balanced' | 'quality',
     systemInstructions: string,
+    input: SearchInput,
+    emitter: EventEmitter,
   ) {
-    return RunnableSequence.from([
+    const chatPrompt = ChatPromptTemplate.fromMessages([
-      RunnableMap.from({
+      ['system', this.config.responsePrompt],
-        systemInstructions: () => systemInstructions,
+      new MessagesPlaceholder('chat_history'),
-        query: (input: BasicChainInput) => input.query,
+      ['user', '{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 = '';
+
+    if (optimizationMode === 'speed' || optimizationMode === 'balanced') {
       let docs: Document[] | null = null;
       let query = input.query;
 
       if (this.config.searchWeb) {
-            const searchRetrieverChain =
+        const searchResults = await this.searchSources(llm, input, emitter);
-              await this.createSearchRetrieverChain(llm);
 
-            const searchRetrieverResult = await searchRetrieverChain.invoke({
+        query = searchResults.query;
-              chat_history: processedHistory,
+        docs = searchResults.docs;
-              query,
-            });
-
-            query = searchRetrieverResult.query;
-            docs = searchRetrieverResult.docs;
       }
 
       const sortedDocs = await this.rerankDocs(
@@ -274,23 +399,42 @@ class MetaSearchAgent implements MetaSearchAgentType {
         optimizationMode,
       );
 
-          return sortedDocs;
+      emitter.emit(
-        })
+        'data',
-          .withConfig({
+        JSON.stringify({ type: 'sources', data: sortedDocs }),
-            runName: 'FinalSourceRetriever',
+      );
-          })
+
-          .pipe(this.processDocs),
+      context = this.processDocs(sortedDocs);
-      }),
+    } else if (optimizationMode === 'quality') {
-      ChatPromptTemplate.fromMessages([
+      let docs: Document[] = [];
-        ['system', this.config.responsePrompt],
+
-        new MessagesPlaceholder('chat_history'),
+      docs = await this.performDeepResearch(llm, input, emitter);
-        ['user', '{query}'],
+
-      ]),
+      emitter.emit('data', JSON.stringify({ type: 'sources', data: docs }));
-      llm,
+
-      this.strParser,
+      context = this.processDocs(docs);
-    ]).withConfig({
+    }
-      runName: 'FinalResponseGenerator',
+
+    const formattedChatPrompt = await chatPrompt.invoke({
+      query: input.query,
+      chat_history: input.chat_history,
+      date: new Date().toISOString(),
+      context: context,
+      systemInstructions: systemInstructions,
     });
+
+    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(
@@ -426,44 +570,13 @@ 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[],
@@ -475,26 +588,19 @@ class MetaSearchAgent implements MetaSearchAgentType {
   ) {
     const emitter = new eventEmitter();
 
-    const answeringChain = await this.createAnsweringChain(
+    this.streamAnswer(
       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;
 }
 
-interface SearxngSearchResult {
+export interface SearxngSearchResult {
   title: string;
   url: string;
   img_src?: string;