@lobehub/chat 1.77.16 → 1.77.18

package/CHANGELOG.md

@@ -2,6 +2,56 @@
 
 # Changelog
 
+### [Version 1.77.18](https://github.com/lobehub/lobe-chat/compare/v1.77.17...v1.77.18)
+
+<sup>Released on **2025-04-09**</sup>
+
+#### 💄 Styles
+
+- **misc**: Add `time_range` & `categories` support for SearXNG.
+
+<br/>
+
+<details>
+<summary><kbd>Improvements and Fixes</kbd></summary>
+
+#### Styles
+
+- **misc**: Add `time_range` & `categories` support for SearXNG, closes [#6813](https://github.com/lobehub/lobe-chat/issues/6813) ([9e4cd8c](https://github.com/lobehub/lobe-chat/commit/9e4cd8c))
+
+</details>
+
+<div align="right">
+
+[![](https://img.shields.io/badge/-BACK_TO_TOP-151515?style=flat-square)](#readme-top)
+
+</div>
+
+### [Version 1.77.17](https://github.com/lobehub/lobe-chat/compare/v1.77.16...v1.77.17)
+
+<sup>Released on **2025-04-08**</sup>
+
+#### 🐛 Bug Fixes
+
+- **misc**: Refactor ollama pull flow and model service.
+
+<br/>
+
+<details>
+<summary><kbd>Improvements and Fixes</kbd></summary>
+
+#### What's fixed
+
+- **misc**: Refactor ollama pull flow and model service, closes [#7330](https://github.com/lobehub/lobe-chat/issues/7330) ([44d63b8](https://github.com/lobehub/lobe-chat/commit/44d63b8))
+
+</details>
+
+<div align="right">
+
+[![](https://img.shields.io/badge/-BACK_TO_TOP-151515?style=flat-square)](#readme-top)
+
+</div>
+
 ### [Version 1.77.16](https://github.com/lobehub/lobe-chat/compare/v1.77.15...v1.77.16)
 
 <sup>Released on **2025-04-06**</sup>

package/changelog/v1.json

@@ -1,4 +1,22 @@
 [
+  {
+    "children": {
+      "improvements": [
+        "Add time_range & categories support for SearXNG."
+      ]
+    },
+    "date": "2025-04-09",
+    "version": "1.77.18"
+  },
+  {
+    "children": {
+      "fixes": [
+        "Refactor ollama pull flow and model service."
+      ]
+    },
+    "date": "2025-04-08",
+    "version": "1.77.17"
+  },
   {
     "children": {
       "improvements": [

package/contributing/Basic/Architecture.md

@@ -1,6 +1,6 @@
 # Architecture Design
 
-LobeChat is an AI conversation application built on the Next.js framework, aiming to provide an AI productivity platform that enables users to interact with AI through natural language. The following is an overview of the architecture design of LobeChat:
+LobeChat is an AI chat application built on the Next.js framework, aiming to provide an AI productivity platform that enables users to interact with AI through natural language. The following is an overview of the architecture design of LobeChat:
 
 #### TOC
 

package/contributing/Basic/Architecture.zh-CN.md

@@ -1,6 +1,6 @@
 # 架构设计
 
-LobeChat 是一个基于 Next.js 框架构建的 AI 会话应用，旨在提供一个 AI 生产力平台，使用户能够与 AI 进行自然语言交互。以下是 LobeChat 的架构设计介稿：
+LobeChat 是一个基于 Next.js 框架构建的 AI 聊天应用，旨在提供一个 AI 生产力平台，使用户能够与 AI 进行自然语言交互。以下是 LobeChat 的架构设计介稿：
 
 #### TOC
 

package/contributing/Basic/Chat-API.md

@@ -1,136 +1,354 @@
-# Conversation API Implementation Logic
+# Lobe Chat API Client-Server Interaction Logic
 
-The implementation of LobeChat's large model AI mainly relies on OpenAI's API, including the core conversation API on the backend and the integrated API on the frontend. Next, we will introduce the implementation approach and code for the backend and frontend separately.
+This document explains the implementation logic of Lobe Chat API in client-server interactions, including event sequences and core components involved.
 
-#### TOC
+## Table of Contents
 
-- [Backend Implementation](#backend-implementation)
-  - [Core Conversation API](#core-conversation-api)
-  - [Conversation Result Processing](#conversation-result-processing)
-- [Frontend Implementation](#frontend-implementation)
-  - [Frontend Integration](#frontend-integration)
-  - [Using Streaming to Get Results](#using-streaming-to-get-results)
+- [Interaction Sequence Diagram](#interaction-sequence-diagram)
+- [Main Process Steps](#main-process-steps)
+- [AgentRuntime Overview](#agentruntime-overview)
 
-## Backend Implementation
+## Interaction Sequence Diagram
 
-The following code removes authentication, error handling, and other logic, retaining only the core functionality logic.
+```mermaid
+sequenceDiagram
+    participant Client as Frontend Client
+    participant ChatService as Frontend ChatService
+    participant ChatAPI as Backend Chat API
+    participant AgentRuntime as AgentRuntime
+    participant ModelProvider as Model Provider API
+    participant PluginGateway as Plugin Gateway
 
-### Core Conversation API
+    Client->>ChatService: Call createAssistantMessage
+    Note over ChatService: Process messages, tools, and parameters
 
-In the file `src/app/api/openai/chat/handler.ts`, we define a `POST` method, which first parses the payload data from the request (i.e., the conversation content sent by the client), and then retrieves the authorization information from the request. Then, we create an `openai` object and call the `createChatCompletion` method, which is responsible for sending the conversation request to OpenAI and returning the result.
+    ChatService->>ChatService: Call getChatCompletion
+    Note over ChatService: Prepare request parameters
 
-```ts
-export const POST = async (req: Request) => {
-  const payload = await req.json();
+    ChatService->>ChatAPI: Send POST request to /webapi/chat/[provider]
 
-  const { apiKey, endpoint } = getOpenAIAuthFromRequest(req);
+    ChatAPI->>AgentRuntime: Initialize AgentRuntime
+    Note over AgentRuntime: Create runtime with provider and user config
 
-  const openai = createOpenai(apiKey, endpoint);
+    ChatAPI->>AgentRuntime: Call chat method
+    AgentRuntime->>ModelProvider: Send chat completion request
 
-  return createChatCompletion({ openai, payload });
-};
-```
+    ModelProvider-->>AgentRuntime: Return streaming response
+    AgentRuntime-->>ChatAPI: Process response and return stream
 
-### Conversation Result Processing
-
-In the file `src/app/api/openai/chat/createChatCompletion.ts`, we define the `createChatCompletion` method, which first preprocesses the payload data, then calls OpenAI's `chat.completions.create` method to send the request, and uses the `OpenAIStream` from the [Vercel AI SDK](https://sdk.vercel.ai/docs) to convert the returned result into a streaming response.
-
-```ts
-import { OpenAIStream, StreamingTextResponse } from 'ai';
-
-export const createChatCompletion = async ({ payload, openai }: CreateChatCompletionOptions) => {
-  const { messages, ...params } = payload;
-
-  const formatMessages = messages.map((m) => ({
-    content: m.content,
-    name: m.name,
-    role: m.role,
-  }));
-
-  const response = await openai.chat.completions.create(
-    {
-      messages: formatMessages,
-      ...params,
-      stream: true,
-    },
-    { headers: { Accept: '*/*' } },
-  );
-  const stream = OpenAIStream(response);
-  return new StreamingTextResponse(stream);
-};
-```
+    ChatAPI-->>ChatService: Stream back SSE response
+
+    ChatService->>ChatService: Handle streaming response with fetchSSE
+    Note over ChatService: Process event stream with fetchEventSource
+
+    loop For each data chunk
+        ChatService->>ChatService: Handle different event types (text, tool_calls, reasoning, etc.)
+        ChatService-->>Client: Return current chunk via onMessageHandle callback
+    end
+
+    ChatService-->>Client: Return complete result via onFinish callback
 
-## Frontend Implementation
-
-### Frontend Integration
-
-In the `src/services/chatModel.ts` file, we define the `fetchChatModel` method, which first preprocesses the payload data, then sends a POST request to the `/chat` endpoint on the backend, and returns the request result.
-
-```ts
-export const fetchChatModel = (
-  { plugins: enabledPlugins, ...params }: Partial<OpenAIStreamPayload>,
-  options?: FetchChatModelOptions,
-) => {
-  const payload = merge(
-    {
-      model: initialLobeAgentConfig.model,
-      stream: true,
-      ...initialLobeAgentConfig.params,
-    },
-    params,
-  );
-
-  const filterFunctions: ChatCompletionFunctions[] = pluginSelectors.enabledSchema(enabledPlugins)(
-    usePluginStore.getState(),
-  );
-
-  const functions = filterFunctions.length === 0 ? undefined : filterFunctions;
-
-  return fetch(OPENAI_URLS.chat, {
-    body: JSON.stringify({ ...payload, functions }),
-    headers: createHeaderWithOpenAI({ 'Content-Type': 'application/json' }),
-    method: 'POST',
-    signal: options?.signal,
-  });
-};
+    Note over ChatService,ModelProvider: Plugin calling scenario
+    ModelProvider-->>ChatService: Return response with tool_calls
+    ChatService->>ChatService: Parse tool calls
+    ChatService->>ChatService: Call runPluginApi
+    ChatService->>PluginGateway: Send plugin request to gateway
+    PluginGateway-->>ChatService: Return plugin execution result
+    ChatService->>ModelProvider: Return plugin result to model
+    ModelProvider-->>ChatService: Generate final response based on plugin result
+
+    Note over ChatService,ModelProvider: Preset task scenario
+    Client->>ChatService: Trigger preset task (e.g., translation, search)
+    ChatService->>ChatService: Call fetchPresetTaskResult
+    ChatService->>ChatAPI: Send preset task request
+    ChatAPI-->>ChatService: Return task result
+    ChatService-->>Client: Return result via callback function
 ```
 
-### Using Streaming to Get Results
+## Main Process Steps
 
-In the `src/utils/fetch.ts` file, we define the `fetchSSE` method, which uses a streaming approach to retrieve data. When a new data chunk is read, it calls the `onMessageHandle` callback function to process the data chunk, achieving a typewriter-like output effect.
+1. **Client Initiates Request**: The client calls the createAssistantMessage method of the frontend ChatService.
 
-```ts
-export const fetchSSE = async (fetchFn: () => Promise<Response>, options: FetchSSEOptions = {}) => {
-  const response = await fetchFn();
+2. **Frontend Processes Request**:
 
-  if (!response.ok) {
-    const chatMessageError = await getMessageError(response);
+   - `src/services/chat.ts` preprocesses messages, tools, and parameters
+   - Calls getChatCompletion to prepare request parameters
+   - Uses `src/utils/fetch/fetchSSE.ts` to send request to backend API
 
-    options.onErrorHandle?.(chatMessageError);
-    return;
-  }
+3. **Backend Processes Request**:
 
-  const returnRes = response.clone();
+   - `src/app/(backend)/webapi/chat/[provider]/route.ts` receives the request
+   - Initializes AgentRuntime
+   - Creates the appropriate model instance based on user configuration and provider
 
-  const data = response.body;
+4. **Model Call**:
 
-  if (!data) return;
+   - `src/libs/agent-runtime/AgentRuntime.ts` calls the respective model provider's API
+   - Returns streaming response
 
-  const reader = data.getReader();
-  const decoder = new TextDecoder();
+5. **Process Response**:
+
+   - Backend converts model response to Stream and returns it
+   - Frontend processes streaming response via fetchSSE and [fetchEventSource](https://github.com/Azure/fetch-event-source)
+   - Handles different types of events (text, tool calls, reasoning, etc.)
+   - Passes results back to client through callback functions
 
-  let done = false;
+6. **Plugin Calling Scenario**:
 
-  while (!done) {
-    const { value, done: doneReading } = await reader.read();
-    done = doneReading;
-    const chunkValue = decoder.decode(value);
+   When the AI model returns a `tool_calls` field in its response, it triggers the plugin calling process:
 
-    options.onMessageHandle?.(chunkValue);
-  }
+   - AI model returns response containing `tool_calls`, indicating a need to call tools
+   - Frontend handles tool calls via the `internal_callPluginApi` method
+   - Calls `runPluginApi` method to execute plugin functionality, including retrieving plugin settings and manifest, creating authentication headers, and sending requests to the plugin gateway
+   - After plugin execution completes, the result is returned to the AI model, which generates the final response based on the result
 
-  return returnRes;
-};
-```
+   **Real-world Examples**:
+
+   - **Search Plugin**: When a user needs real-time information, the AI calls a web search plugin to retrieve the latest data
+   - **DALL-E Plugin**: When a user requests image generation, the AI calls the DALL-E plugin to create images
+   - **Midjourney Plugin**: Provides higher quality image generation capabilities by calling the Midjourney service via API
+
+7. **Preset Task Processing**:
+
+   Preset tasks are specific predefined functions that are typically triggered when users perform specific actions (rather than being part of the regular chat flow). These tasks use the `fetchPresetTaskResult` method, which is similar to the normal chat flow but uses specially designed prompt chains.
+
+   **Execution Timing**: Preset tasks are mainly triggered in the following scenarios:
+
+   1. **Agent Information Auto-generation**: Triggered when users create or edit an agent
+
+      - Agent avatar generation (via `autoPickEmoji` method)
+      - Agent description generation (via `autocompleteAgentDescription` method)
+      - Agent tag generation (via `autocompleteAgentTags` method)
+      - Agent title generation (via `autocompleteAgentTitle` method)
+
+   2. **Message Translation**: Triggered when users manually click the translate button (via `translateMessage` method)
+
+   3. **Web Search**: When search is enabled but the model doesn't support tool calling, search functionality is implemented via `fetchPresetTaskResult`
+
+   **Code Examples**:
+
+   Agent avatar auto-generation implementation:
 
-The above is the core implementation of the LobeChat session API. With an understanding of these core codes, further expansion and optimization of LobeChat's AI functionality can be achieved.
+   ```typescript
+   // src/features/AgentSetting/store/action.ts
+   autoPickEmoji: async () => {
+     const { config, meta, dispatchMeta } = get();
+     const systemRole = config.systemRole;
+
+     chatService.fetchPresetTaskResult({
+       onFinish: async (emoji) => {
+         dispatchMeta({ type: 'update', value: { avatar: emoji } });
+       },
+       onLoadingChange: (loading) => {
+         get().updateLoadingState('avatar', loading);
+       },
+       params: merge(
+         get().internal_getSystemAgentForMeta(),
+         chainPickEmoji([meta.title, meta.description, systemRole].filter(Boolean).join(',')),
+       ),
+       trace: get().getCurrentTracePayload({ traceName: TraceNameMap.EmojiPicker }),
+     });
+   };
+   ```
+
+   Translation feature implementation:
+
+   ```typescript
+   // src/store/chat/slices/translate/action.ts
+   translateMessage: async (id, targetLang) => {
+     // ...omitted code...
+
+     // Detect language
+     chatService.fetchPresetTaskResult({
+       onFinish: async (data) => {
+         if (data && supportLocales.includes(data)) from = data;
+         await updateMessageTranslate(id, { content, from, to: targetLang });
+       },
+       params: merge(translationSetting, chainLangDetect(message.content)),
+       trace: get().getCurrentTracePayload({ traceName: TraceNameMap.LanguageDetect }),
+     });
+
+     // Perform translation
+     chatService.fetchPresetTaskResult({
+       onMessageHandle: (chunk) => {
+         if (chunk.type === 'text') {
+           content = chunk.text;
+           internal_dispatchMessage({
+             id,
+             type: 'updateMessageTranslate',
+             value: { content, from, to: targetLang },
+           });
+         }
+       },
+       onFinish: async () => {
+         await updateMessageTranslate(id, { content, from, to: targetLang });
+         internal_toggleChatLoading(false, id, n('translateMessage(end)', { id }) as string);
+       },
+       params: merge(translationSetting, chainTranslate(message.content, targetLang)),
+       trace: get().getCurrentTracePayload({ traceName: TraceNameMap.Translation }),
+     });
+   };
+   ```
+
+8. **Completion**:
+   - When the stream ends, the onFinish callback is called, providing the complete response result
+
+## AgentRuntime Overview
+
+AgentRuntime is a core abstraction layer in Lobe Chat that encapsulates a unified interface for interacting with different AI model providers. Its main responsibilities and features include:
+
+1. **Unified Abstraction Layer**: AgentRuntime provides a unified interface that hides the implementation details and differences between various AI provider APIs (such as OpenAI, Anthropic, Bedrock, etc.).
+
+2. **Model Initialization**: Through the static `initializeWithProvider` method, it initializes the corresponding runtime instance based on the specified provider and configuration parameters.
+
+3. **Capability Encapsulation**:
+
+   - `chat` method: Handles chat streaming requests
+   - `models` method: Retrieves model lists
+   - Supports text embedding, text-to-image, text-to-speech, and other functionalities (if supported by the model provider)
+
+4. **Plugin Architecture**: Through the `src/libs/agent-runtime/runtimeMap.ts` mapping table, it implements an extensible plugin architecture, making it easy to add new model providers. Currently, it supports over 40 different model providers:
+
+   ```typescript
+   export const providerRuntimeMap = {
+     openai: LobeOpenAI,
+     anthropic: LobeAnthropicAI,
+     google: LobeGoogleAI,
+     azure: LobeAzureOpenAI,
+     bedrock: LobeBedrockAI,
+     ollama: LobeOllamaAI,
+     // ...over 40 other model providers
+   };
+   ```
+
+5. **Adapter Pattern**: Internally, it uses the adapter pattern to adapt different provider APIs to the unified `src/libs/agent-runtime/BaseAI.ts` interface:
+
+   ```typescript
+   export interface LobeRuntimeAI {
+     baseURL?: string;
+     chat(payload: ChatStreamPayload, options?: ChatCompetitionOptions): Promise<Response>;
+     embeddings?(payload: EmbeddingsPayload, options?: EmbeddingsOptions): Promise<Embeddings[]>;
+     models?(): Promise<any>;
+     textToImage?: (payload: TextToImagePayload) => Promise<string[]>;
+     textToSpeech?: (
+       payload: TextToSpeechPayload,
+       options?: TextToSpeechOptions,
+     ) => Promise<ArrayBuffer>;
+   }
+   ```
+
+   **Adapter Implementation Examples**:
+
+   1. **OpenRouter Adapter**:
+      OpenRouter is a unified API that allows access to AI models from multiple providers. Lobe Chat implements support for OpenRouter through an adapter:
+
+      ```typescript
+      // OpenRouter adapter implementation
+      class LobeOpenRouterAI implements LobeRuntimeAI {
+        client: OpenAI;
+        baseURL: string;
+
+        constructor(options: OpenAICompatibleOptions) {
+          // Initialize OpenRouter client using OpenAI-compatible API
+          this.client = new OpenAI({
+            apiKey: options.apiKey,
+            baseURL: OPENROUTER_BASE_URL,
+            defaultHeaders: {
+              'HTTP-Referer': 'https://github.com/lobehub/lobe-chat',
+              'X-Title': 'LobeChat',
+            },
+          });
+          this.baseURL = OPENROUTER_BASE_URL;
+        }
+
+        // Implement chat functionality
+        async chat(payload: ChatCompletionCreateParamsBase, options?: RequestOptions) {
+          // Convert Lobe Chat request format to OpenRouter format
+          // Handle model mapping, message format, etc.
+          return this.client.chat.completions.create(
+            {
+              ...payload,
+              model: payload.model || 'openai/gpt-4-turbo', // Default model
+            },
+            options,
+          );
+        }
+
+        // Implement other LobeRuntimeAI interface methods
+      }
+      ```
+
+   2. **Google Gemini Adapter**:
+      Gemini is Google's large language model. Lobe Chat supports Gemini series models through a dedicated adapter:
+
+      ```typescript
+      import { GoogleGenerativeAI } from '@google/generative-ai';
+
+      // Gemini adapter implementation
+      class LobeGoogleAI implements LobeRuntimeAI {
+        client: GoogleGenerativeAI;
+        baseURL: string;
+        apiKey: string;
+
+        constructor(options: GoogleAIOptions) {
+          // Initialize Google Generative AI client
+          this.client = new GoogleGenerativeAI(options.apiKey);
+          this.apiKey = options.apiKey;
+          this.baseURL = options.baseURL || GOOGLE_AI_BASE_URL;
+        }
+
+        // Implement chat functionality
+        async chat(payload: ChatCompletionCreateParamsBase, options?: RequestOptions) {
+          // Select appropriate model (supports Gemini Pro, Gemini Flash, etc.)
+          const modelName = payload.model || 'gemini-pro';
+          const model = this.client.getGenerativeModel({ model: modelName });
+
+          // Process multimodal inputs (e.g., images)
+          const contents = this.processMessages(payload.messages);
+
+          // Set generation parameters
+          const generationConfig = {
+            temperature: payload.temperature,
+            topK: payload.top_k,
+            topP: payload.top_p,
+            maxOutputTokens: payload.max_tokens,
+          };
+
+          // Create chat session and get response
+          const chat = model.startChat({
+            generationConfig,
+            history: contents.slice(0, -1),
+            safetySettings: this.getSafetySettings(payload),
+          });
+
+          // Handle streaming response
+          return this.handleStreamResponse(chat, contents, options?.signal);
+        }
+
+        // Implement other processing methods
+        private processMessages(messages) {
+          /* ... */
+        }
+        private getSafetySettings(payload) {
+          /* ... */
+        }
+        private handleStreamResponse(chat, contents, signal) {
+          /* ... */
+        }
+      }
+      ```
+
+   **Different Model Implementations**:
+
+   - `src/libs/agent-runtime/openai/index.ts` - OpenAI implementation
+   - `src/libs/agent-runtime/anthropic/index.ts` - Anthropic implementation
+   - `src/libs/agent-runtime/google/index.ts` - Google implementation
+   - `src/libs/agent-runtime/openrouter/index.ts` - OpenRouter implementation
+
+For detailed implementation, see:
+
+- `src/libs/agent-runtime/AgentRuntime.ts` - Core runtime class
+- `src/libs/agent-runtime/BaseAI.ts` - Define base interface
+- `src/libs/agent-runtime/runtimeMap.ts` - Provider mapping table
+- `src/libs/agent-runtime/UniformRuntime/index.ts` - Handle multi-model unified runtime
+- `src/libs/agent-runtime/utils/openaiCompatibleFactory/index.ts` - OpenAI compatible adapter factory