Merge 64136b8410 into 4154d5e4b1

Signed-off-by: Navratan Lal Gupta <navilg0409@gmail.com>

README.md

@@ -84,16 +84,18 @@ There are mainly 2 ways of installing Perplexica - With Docker, Without Docker.
 
 3. After cloning, navigate to the directory containing the project files.
 
-4. Rename the `sample.config.toml` file to `config.toml`. For Docker setups, you need only fill in the following fields:
+4. Update environment variables in `docker-compose.yml` file to configure `config.toml`. 
 
-   - `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**.
-   - `ANTHROPIC`: Your Anthropic API key. **You only need to fill this if you wish to use Anthropic models**.
+Example: 
 
-     **Note**: You can change these after starting Perplexica from the settings dialog.
+Below section in `config.toml` can be configured using variables `MODELS_CUSTOM_OPENAI_API_KEY="sk-123456"`, `MODELS_CUSTOM_OPENAI_API_URL="http://localopenai:11134"` and `MODELS_CUSTOM_OPENAI_MODEL_NAME="meta-llama/llama-4"`
 
-   - `SIMILARITY_MEASURE`: The similarity measure to use (This is filled by default; you can leave it as is if you are unsure about it.)
+```toml
+[MODELS.CUSTOM_OPENAI]
+API_KEY = "sk-123456"
+API_URL = "http://localopenai:11134"
+MODEL_NAME = "meta-llama/llama-4"
+```
 
 5. Ensure you are in the directory containing the `docker-compose.yaml` file and execute:
 
@@ -159,6 +161,7 @@ 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
 

app.dockerfile

@@ -21,7 +21,9 @@ COPY --from=builder /home/perplexica/.next/static ./public/_next/static
 
 COPY --from=builder /home/perplexica/.next/standalone ./
 COPY --from=builder /home/perplexica/data ./data
+COPY sample.config.toml /home/perplexica/config.toml
+COPY container_entrypoint.sh /home/perplexica/container_entrypoint.sh
 
 RUN mkdir /home/perplexica/uploads
 
-CMD ["node", "server.js"]
+CMD ["bash", "/home/perplexica/container_entrypoint.sh"]

container_entrypoint.sh

@@ -0,0 +1,49 @@
+#!/usr/bin/env bash
+
+CONFIG_TOML_FILE=/home/perplexica/config.toml
+
+TMP_FILE=${CONFIG_TOML_FILE}.tmp
+touch $TMP_FILE
+
+while IFS= read -r line; do
+    # Check if line is a section header (e.g., "[GENERAL]")
+    if [[ "$line" =~ ^\[([^]]+)\] ]]; then
+        current_section="${BASH_REMATCH[1]}"
+        echo "$line" >> "$TMP_FILE"
+        continue
+    fi
+
+    # Skip empty lines and comments
+    if [[ -z "$line" || "$line" =~ ^[[:space:]]*\# ]]; then
+        echo "$line" >> "$TMP_FILE"
+        continue
+    fi
+
+    # Extract key and value (handling quoted values)
+    key=$(echo "$line" | cut -d '=' -f 1 | xargs)
+    value=$(echo "$line" | cut -d '=' -f 2- | xargs)
+
+
+    # Construct the environment variable name in form of SECTION_KEY (e.g., GENERAL_SIMILARITY_MEASURE, MODELS_GEMINI_API_KEY)
+    current_section=$(echo "$current_section" | sed 's/\./_/')
+    env_var_name="${current_section}_${key}"
+
+    # Check if the environment variable exists
+    env_var_value=$(echo "${!env_var_name}")
+    if [ -n "$env_var_value" ]; then
+        new_value="$env_var_value"
+        echo "$key = $new_value" >> "$TMP_FILE"
+    else
+        # Keep original line if no env var exists
+        echo "$line" >> "$TMP_FILE"
+    fi
+
+done < "$CONFIG_TOML_FILE"
+
+# Replace the original file
+mv "$TMP_FILE" "$CONFIG_TOML_FILE"
+
+echo "Config file updated successfully."
+
+# Start server
+node server.js

docker-compose.yaml

@@ -16,6 +16,19 @@ services:
       dockerfile: app.dockerfile
     environment:
       - SEARXNG_API_URL=http://searxng:8080
+      - GENERAL_SIMILARITY_MEASURE="cosine" # "cosine" or "dot"
+      - GENERAL_KEEP_ALIVE="5m"             # How long to keep Ollama models loaded into memory. (Instead of using -1 use "-1m")
+      - MODELS_OPENAI_API_KEY=""
+      - MODELS_GROQ_API_KEY=""
+      - MODELS_ANTHROPIC_API_KEY=""
+      - MODELS_GEMINI_API_KEY=""
+      - MODELS_CUSTOM_OPENAI_API_KEY=""
+      - MODELS_CUSTOM_OPENAI_API_URL=""
+      - MODELS_CUSTOM_OPENAI_MODEL_NAME=""
+      - MODELS_OLLAMA_API_KEY=""            # Ollama API URL - http://host.docker.internal:11434
+      - MODELS_DEEPSEEK_API_KEY=""
+      - MODELS_LM_STUDIO_API_KEY=""         # LM Studio API URL - http://host.docker.internal:1234
+      - API_ENDPOINTS_SEARXNG=""            # SearxNG API URL - http://localhost:32768
     ports:
       - 3000:3000
     networks:
@@ -23,7 +36,6 @@ services:
     volumes:
       - backend-dbstore:/home/perplexica/data
       - uploads:/home/perplexica/uploads
-      - ./config.toml:/home/perplexica/config.toml
     restart: unless-stopped
 
 networks:

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

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/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: {
@@ -20,6 +21,7 @@ interface SettingsType {
   anthropicApiKey: string;
   geminiApiKey: string;
   ollamaApiUrl: string;
+  lmStudioApiUrl: string;
   deepseekApiKey: string;
   customOpenaiApiKey: string;
   customOpenaiApiUrl: string;
@@ -548,6 +550,7 @@ const Page = () => {
                         (provider) => ({
                           value: provider,
                           label:
+                            (PROVIDER_METADATA as any)[provider]?.displayName ||
                             provider.charAt(0).toUpperCase() +
                               provider.slice(1),
                         }),
@@ -690,6 +693,7 @@ const Page = () => {
                         (provider) => ({
                           value: provider,
                           label:
+                            (PROVIDER_METADATA as any)[provider]?.displayName ||
                             provider.charAt(0).toUpperCase() +
                               provider.slice(1),
                         }),
@@ -858,6 +862,25 @@ 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/ChatWindow.tsx

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

src/components/MessageBox.tsx

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

@@ -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(
+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>) => {
+  // 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

@@ -1,6 +1,11 @@
 import axios from 'axios';
 import { getKeepAlive, getOllamaApiEndpoint } from '../config';
 import { ChatModel, EmbeddingModel } from '.';
+
+export const PROVIDER_INFO = {
+  key: 'ollama',
+  displayName: 'Ollama',
+};
 import { ChatOllama } from '@langchain/community/chat_models/ollama';
 import { OllamaEmbeddings } from '@langchain/community/embeddings/ollama';
 

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;
 };
@@ -56,25 +60,14 @@ 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,
-    );
-
-    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);
-
+    return RunnableSequence.from([
+      PromptTemplate.fromTemplate(this.config.queryGeneratorPrompt),
+      llm,
+      this.strParser,
+      RunnableLambda.from(async (input: string) => {
         const linksOutputParser = new LineListOutputParser({
           key: 'links',
         });
@@ -83,10 +76,10 @@ class MetaSearchAgent implements MetaSearchAgentType {
           key: 'question',
         });
 
-    const links = await linksOutputParser.parse(messageStr);
+        const links = await linksOutputParser.parse(input);
         let question = this.config.summarizer
-      ? await questionOutputParser.parse(messageStr)
-      : messageStr;
+          ? await questionOutputParser.parse(input)
+          : input;
 
         if (question === 'not_needed') {
           return { query: '', docs: [] };
@@ -106,7 +99,8 @@ class MetaSearchAgent implements MetaSearchAgentType {
           linkDocs.map((doc) => {
             const URLDocExists = docGroups.find(
               (d) =>
-            d.metadata.url === doc.metadata.url && d.metadata.totalDocs < 10,
+                d.metadata.url === doc.metadata.url &&
+                d.metadata.totalDocs < 10,
             );
 
             if (!URLDocExists) {
@@ -121,7 +115,8 @@ class MetaSearchAgent implements MetaSearchAgentType {
 
             const docIndex = docGroups.findIndex(
               (d) =>
-            d.metadata.url === doc.metadata.url && d.metadata.totalDocs < 10,
+                d.metadata.url === doc.metadata.url &&
+                d.metadata.totalDocs < 10,
             );
 
             if (docIndex !== -1) {
@@ -233,162 +228,42 @@ class MetaSearchAgent implements MetaSearchAgentType {
 
           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 = '';
-
-    if (optimizationMode === 'speed' || optimizationMode === 'balanced') {
           let docs: Document[] | null = null;
           let query = input.query;
 
           if (this.config.searchWeb) {
-        const searchResults = await this.searchSources(llm, input, emitter);
+            const searchRetrieverChain =
+              await this.createSearchRetrieverChain(llm);
 
-        query = searchResults.query;
-        docs = searchResults.docs;
+            const searchRetrieverResult = await searchRetrieverChain.invoke({
+              chat_history: processedHistory,
+              query,
+            });
+
+            query = searchRetrieverResult.query;
+            docs = searchRetrieverResult.docs;
           }
 
           const sortedDocs = await this.rerankDocs(
@@ -399,42 +274,23 @@ class MetaSearchAgent implements MetaSearchAgentType {
             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;

src/lib/utils/documents.ts

@@ -64,7 +64,7 @@ export const getDocumentsFromLinks = async ({ links }: { links: string[] }) => {
         const splittedText = await splitter.splitText(parsedText);
         const title = res.data
           .toString('utf8')
-          .match(/<title>(.*?)<\/title>/)?.[1];
+          .match(/<title.*>(.*?)<\/title>/)?.[1];
 
         const linkDocs = splittedText.map((text) => {
           return new Document({