✨ feat: 优化插件模式下的用户体验 (#13)

next.config.mjs

@@ -13,7 +13,7 @@ const withPWA = nextPWA({
 const nextConfig = {
   reactStrictMode: true,
   pageExtensions: ['page.tsx', 'api.ts'],
-  transpilePackages: ['@lobehub/ui', 'antd-style'],
+  transpilePackages: ['@lobehub/ui'],
 
   webpack(config) {
     config.experiments = {
@@ -31,8 +31,8 @@ const nextConfig = {
         destination: `${API_END_PORT_URL}/api/openai`,
       },
       {
-        source: '/api/chain-dev',
-        destination: `${API_END_PORT_URL}/api/chain`,
+        source: '/api/plugins-dev',
+        destination: `${API_END_PORT_URL}/api/plugins`,
       },
     ];
   },

src/const/message.ts

@@ -0,0 +1,2 @@
+// 只要 start with 这个，就可以判断为 function message
+export const FUNCTION_MESSAGE_FLAG = '{"function';

src/locales/default/common.ts

@@ -38,6 +38,7 @@ export default {
   'plugin-realtimeWeather': '实时天气预报',
   'plugin-searchEngine': '搜索引擎',
   'pluginList': '插件列表',
+  'pluginLoading': '插件运行中...',
   'profile': '助手身份',
   'reset': '重置',
   'searchAgentPlaceholder': '搜索助手和对话...',

src/pages/api/openai.api.ts

@@ -1,74 +1,15 @@
-import { OpenAIStream, StreamingTextResponse } from 'ai';
-import { Configuration, OpenAIApi } from 'openai-edge';
-import { ChatCompletionFunctions, ChatCompletionRequestMessage } from 'openai-edge/types/api';
+import { StreamingTextResponse } from 'ai';
 
 import { OpenAIStreamPayload } from '@/types/openai';
 
-import pluginList from '../../plugins';
+import { createChatCompletion } from './openai';
 
 export const runtime = 'edge';
 
-const isDev = process.env.NODE_ENV === 'development';
-const OPENAI_PROXY_URL = process.env.OPENAI_PROXY_URL;
-
-// Create an OpenAI API client (that's edge friendly!)
-const config = new Configuration({
-  apiKey: process.env.OPENAI_API_KEY,
-});
-
-const openai = new OpenAIApi(config, isDev && OPENAI_PROXY_URL ? OPENAI_PROXY_URL : undefined);
-
 export default async function handler(req: Request) {
-  const {
-    messages,
-    plugins: enabledPlugins,
-    ...params
-  } = (await req.json()) as OpenAIStreamPayload;
+  const payload = (await req.json()) as OpenAIStreamPayload;
 
-  // ============  1. 前置处理 functions   ============ //
-
-  const filterFunctions: ChatCompletionFunctions[] = pluginList
-    .filter((p) => {
-      // 如果不存在 enabledPlugins，那么全部不启用
-      if (!enabledPlugins) return false;
-
-      // 如果存在 enabledPlugins，那么只启用 enabledPlugins 中的插件
-      return enabledPlugins.includes(p.name);
-    })
-    .map((f) => f.schema);
-
-  const functions = filterFunctions.length === 0 ? undefined : filterFunctions;
-
-  // ============  2. 前置处理 messages   ============ //
-  const formatMessages = messages.map((m) => ({ content: m.content, role: m.role }));
-
-  const response = await openai.createChatCompletion({
-    functions,
-    messages: formatMessages,
-    stream: true,
-    ...params,
-  });
-
-  const stream = OpenAIStream(response, {
-    experimental_onFunctionCall: async ({ name, arguments: args }, createFunctionCallMessages) => {
-      console.log(`执行 functionCall [${name}]`, 'args:', args);
-
-      const func = pluginList.find((f) => f.name === name);
-
-      if (func) {
-        const result = await func.runner(args as any);
-
-        const newMessages = createFunctionCallMessages(result) as ChatCompletionRequestMessage[];
-
-        return openai.createChatCompletion({
-          functions,
-          messages: [...formatMessages, ...newMessages],
-          stream: true,
-          ...params,
-        });
-      }
-    },
-  });
+  const stream = await createChatCompletion(payload);
 
   return new StreamingTextResponse(stream);
 }

src/pages/api/openai.ts

@@ -0,0 +1,50 @@
+import { OpenAIStream, OpenAIStreamCallbacks } from 'ai';
+import { Configuration, OpenAIApi } from 'openai-edge';
+import { ChatCompletionFunctions } from 'openai-edge/types/api';
+
+import { OpenAIStreamPayload } from '@/types/openai';
+
+import pluginList from '../../plugins';
+
+const isDev = process.env.NODE_ENV === 'development';
+const OPENAI_PROXY_URL = process.env.OPENAI_PROXY_URL;
+
+// Create an OpenAI API client (that's edge friendly!)
+const config = new Configuration({
+  apiKey: process.env.OPENAI_API_KEY,
+});
+
+export const openai = new OpenAIApi(
+  config,
+  isDev && OPENAI_PROXY_URL ? OPENAI_PROXY_URL : undefined,
+);
+
+export const createChatCompletion = async (
+  payload: OpenAIStreamPayload,
+  callbacks?: (payload: OpenAIStreamPayload) => OpenAIStreamCallbacks,
+) => {
+  const { messages, plugins: enabledPlugins, ...params } = payload;
+
+  // ============  1. 前置处理 functions   ============ //
+
+  const filterFunctions: ChatCompletionFunctions[] = pluginList
+    .filter((p) => {
+      // 如果不存在 enabledPlugins，那么全部不启用
+      if (!enabledPlugins) return false;
+
+      // 如果存在 enabledPlugins，那么只启用 enabledPlugins 中的插件
+      return enabledPlugins.includes(p.name);
+    })
+    .map((f) => f.schema);
+
+  const functions = filterFunctions.length === 0 ? undefined : filterFunctions;
+
+  // ============  2. 前置处理 messages   ============ //
+  const formatMessages = messages.map((m) => ({ content: m.content, role: m.role }));
+
+  const requestParams = { functions, messages: formatMessages, stream: true, ...params };
+
+  const response = await openai.createChatCompletion(requestParams);
+
+  return OpenAIStream(response, callbacks?.(requestParams));
+};

src/pages/api/plugins.api.ts

@@ -0,0 +1,34 @@
+import { StreamingTextResponse } from 'ai';
+import { ChatCompletionRequestMessage } from 'openai-edge';
+
+import { OpenAIStreamPayload } from '@/types/openai';
+
+import pluginList from '../../plugins';
+import { createChatCompletion, openai } from './openai';
+
+export const runtime = 'edge';
+
+export default async function handler(req: Request) {
+  const payload = (await req.json()) as OpenAIStreamPayload;
+
+  const stream = await createChatCompletion(payload, (payload) => ({
+    experimental_onFunctionCall: async ({ name, arguments: args }, createFunctionCallMessages) => {
+      console.log(`执行 functionCall [${name}]`, 'args:', args);
+
+      const func = pluginList.find((f) => f.name === name);
+
+      if (func) {
+        const result = await func.runner(args as any);
+
+        const newMessages = createFunctionCallMessages(result) as ChatCompletionRequestMessage[];
+
+        return openai.createChatCompletion({
+          ...payload,
+          messages: [...payload.messages, ...newMessages],
+        });
+      }
+    },
+  }));
+
+  return new StreamingTextResponse(stream);
+}

src/pages/chat/[id]/Conversation/ChatList.tsx

@@ -1,10 +1,12 @@
-import { ChatList } from '@lobehub/ui';
+import { ChatList, ChatMessage } from '@lobehub/ui';
 import isEqual from 'fast-deep-equal';
-import { memo } from 'react';
+import { ReactNode, memo } from 'react';
 import { shallow } from 'zustand/shallow';
 
 import { chatSelectors, useSessionStore } from '@/store/session';
+import { isFunctionMessage } from '@/utils/message';
 
+import FunctionMessage from './FunctionMessage';
 import MessageExtra from './MessageExtra';
 
 const List = () => {
@@ -14,6 +16,13 @@ const List = () => {
     shallow,
   );
 
+  const renderMessage = (content: ReactNode, message: ChatMessage) => {
+    if (message.role === 'function')
+      return isFunctionMessage(message.content) ? <FunctionMessage /> : content;
+
+    return content;
+  };
+
   return (
     <ChatList
       data={data}
@@ -33,6 +42,7 @@ const List = () => {
       onMessageChange={(id, content) => {
         dispatchMessage({ id, key: 'content', type: 'updateMessage', value: content });
       }}
+      renderMessage={renderMessage}
       renderMessageExtra={MessageExtra}
       style={{ marginTop: 24 }}
     />

src/pages/chat/[id]/Conversation/FunctionMessage.tsx

@@ -0,0 +1,31 @@
+import { LoadingOutlined } from '@ant-design/icons';
+import { createStyles } from 'antd-style';
+import { memo } from 'react';
+import { useTranslation } from 'react-i18next';
+import { Flexbox } from 'react-layout-kit';
+
+const useStyles = createStyles(({ css, token }) => ({
+  container: css`
+    padding: 4px 8px;
+
+    color: ${token.colorTextSecondary};
+
+    background: ${token.colorFillTertiary};
+    border: 1px solid ${token.colorBorder};
+    border-radius: 6px;
+  `,
+}));
+const FunctionMessage = memo(() => {
+  const { t } = useTranslation();
+  const { styles } = useStyles();
+  return (
+    <Flexbox className={styles.container} gap={8} horizontal>
+      <div>
+        <LoadingOutlined />
+      </div>
+      {t('pluginLoading')}
+    </Flexbox>
+  );
+});
+
+export default FunctionMessage;

src/pages/chat/[id]/Conversation/MessageExtra.tsx

@@ -1,9 +1,13 @@
+import { Avatar, Icon, Tooltip } from '@lobehub/ui';
 import { Tag } from 'antd';
 import { createStyles } from 'antd-style';
+import { LucideToyBrick } from 'lucide-react';
 import { ReactNode } from 'react';
+import { useTranslation } from 'react-i18next';
 import { Flexbox } from 'react-layout-kit';
 import { shallow } from 'zustand/shallow';
 
+import pluginList from '@/plugins';
 import { agentSelectors, useSessionStore } from '@/store/session';
 import { ChatMessage } from '@/types/chatMessage';
 
@@ -11,27 +15,70 @@ const useStyles = createStyles(({ css }) => ({
   container: css`
     margin-top: 8px;
   `,
+  plugin: css`
+    display: flex;
+    gap: 4px;
+    align-items: center;
+    width: fit-content;
+  `,
 }));
 
-const MessageExtra = ({ role, extra }: ChatMessage): ReactNode => {
+const MessageExtra = ({ role, extra, function_call }: ChatMessage): ReactNode => {
   const { styles } = useStyles();
 
+  const { t } = useTranslation();
   const [model] = useSessionStore((s) => [agentSelectors.currentAgentModel(s)], shallow);
 
-  // 1. 只有 ai 的 message
-  // 2. 且存在 fromModel
-  // 3. 且当前的 model 和 fromModel 不一致时
-  if (role === 'assistant' && extra?.fromModel && model !== extra?.fromModel)
-    // 才需要展示 model tag
-    return (
-      <Flexbox className={styles.container}>
-        <div>
-          <Tag bordered={false} style={{ borderRadius: 6 }}>
-            {extra?.fromModel}
-          </Tag>
-        </div>
-      </Flexbox>
-    );
+  const plugin = pluginList.find((p) => p.name === function_call?.name);
+  const funcTag = (
+    <Tooltip title={function_call?.arguments}>
+      <Tag bordered={false} className={styles.plugin} style={{ borderRadius: 6 }}>
+        {plugin?.avatar ? (
+          <Avatar avatar={plugin?.avatar} size={18} />
+        ) : (
+          <Icon icon={LucideToyBrick} />
+        )}
+        {t(`plugin-${function_call?.name}` as any)}
+      </Tag>
+    </Tooltip>
+  );
+
+  const modelTag = (
+    <div>
+      <Tag bordered={false} style={{ borderRadius: 6 }}>
+        {extra?.fromModel}
+      </Tag>
+    </div>
+  );
+
+  // 1. 存在 fromModel
+  // 2. 且当前的 model 和 fromModel 不一致时
+  const hasModelTag = extra?.fromModel && model !== extra?.fromModel;
+
+  const hasFuncTag = !!function_call;
+
+  switch (role) {
+    case 'user':
+    case 'system': {
+      return;
+    }
+    case 'assistant': {
+      // 1. 只有 ai 的 message
+      // 2. 且存在 fromModel
+      // 3. 且当前的 model 和 fromModel 不一致时
+      if (!(hasModelTag || hasFuncTag)) return;
+
+      return (
+        <Flexbox className={styles.container}>
+          {hasFuncTag && funcTag}
+          {hasModelTag && modelTag}
+        </Flexbox>
+      );
+    }
+    case 'function': {
+      return <Flexbox className={styles.container}>{funcTag}</Flexbox>;
+    }
+  }
 };
 
 export default MessageExtra;

src/services/chatModel.ts

@@ -9,7 +9,7 @@ import { URLS } from './url';
  */
 export const fetchChatModel = (
   params: Partial<OpenAIStreamPayload>,
-  signal?: AbortSignal | undefined,
+  options?: { signal?: AbortSignal | undefined; withPlugin?: boolean },
 ) => {
   const payload = merge(
     {
@@ -23,12 +23,12 @@ export const fetchChatModel = (
     params,
   );
 
-  return fetch(URLS.openai, {
+  return fetch(options?.withPlugin ? URLS.plugins : URLS.openai, {
     body: JSON.stringify(payload),
     headers: {
       'Content-Type': 'application/json',
     },
     method: 'POST',
-    signal,
+    signal: options?.signal,
   });
 };

src/services/url.ts

@@ -4,4 +4,5 @@ const prefix = isDev ? '-dev' : '';
 
 export const URLS = {
   openai: '/api/openai' + prefix,
+  plugins: '/api/plugins' + prefix,
 };

src/store/session/slices/chat/action.ts

@@ -3,7 +3,8 @@ import { StateCreator } from 'zustand/vanilla';
 import { fetchChatModel } from '@/services/chatModel';
 import { SessionStore, agentSelectors, chatSelectors, sessionSelectors } from '@/store/session';
 import { ChatMessage } from '@/types/chatMessage';
-import { FetchSSEOptions, fetchSSE } from '@/utils/fetch';
+import { fetchSSE } from '@/utils/fetch';
+import { isFunctionMessage } from '@/utils/message';
 import { nanoid } from '@/utils/uuid';
 
 import { MessageDispatch, messagesReducer } from './messageReducer';
@@ -38,7 +39,11 @@ export interface ChatAction {
    * @param messages - 聊天消息数组
    * @param options - 获取 SSE 选项
    */
-  generateMessage: (messages: ChatMessage[], options: FetchSSEOptions) => Promise<void>;
+  generateMessage: (
+    messages: ChatMessage[],
+    assistantMessageId: string,
+    withPlugin?: boolean,
+  ) => Promise<{ isFunctionCall: boolean; output: string }>;
   /**
    * 处理消息编辑
    * @param messageId - 消息 ID，可选
@@ -100,16 +105,57 @@ export const createChatSlice: StateCreator<
     get().dispatchSession({ chats, id: activeId, type: 'updateSessionChat' });
   },
 
-  generateMessage: async (messages, options) => {
+  generateMessage: async (messages, assistantId, withPlugin) => {
+    const { dispatchMessage } = get();
     set({ chatLoading: true });
     const config = agentSelectors.currentAgentConfigSafe(get());
 
     const fetcher = () =>
-      fetchChatModel({ messages, model: config.model, ...config.params, plugins: config.plugins });
+      fetchChatModel(
+        { messages, model: config.model, ...config.params, plugins: config.plugins },
+        { withPlugin },
+      );
 
-    await fetchSSE(fetcher, options);
+    let output = '';
+    let isFunctionCall = false;
+
+    await fetchSSE(fetcher, {
+      onErrorHandle: (error) => {
+        dispatchMessage({ id: assistantId, key: 'error', type: 'updateMessage', value: error });
+      },
+      onMessageHandle: (text) => {
+        output += text;
+
+        dispatchMessage({
+          id: assistantId,
+          key: 'content',
+          type: 'updateMessage',
+          value: output,
+        });
+
+        // 如果是 function call
+        if (isFunctionMessage(output)) {
+          isFunctionCall = true;
+          // 设为 function
+          dispatchMessage({
+            id: assistantId,
+            key: 'role',
+            type: 'updateMessage',
+            value: 'function',
+          });
+        }
+
+        // 滚动到最后一条消息
+        const item = document.querySelector('#for-loading');
+        if (!item) return;
+
+        item.scrollIntoView({ behavior: 'smooth' });
+      },
+    });
 
     set({ chatLoading: false });
+
+    return { isFunctionCall, output };
   },
 
   handleMessageEditing: (messageId) => {
@@ -145,29 +191,36 @@ export const createChatSlice: StateCreator<
       value: model,
     });
 
-    let output = '';
     // 生成 ai message
-    await generateMessage(messages, {
-      onErrorHandle: (error) => {
-        dispatchMessage({ id: assistantId, key: 'error', type: 'updateMessage', value: error });
-      },
-      onMessageHandle: (text) => {
-        output += text;
+    const { output, isFunctionCall } = await generateMessage(messages, assistantId);
 
-        dispatchMessage({
-          id: assistantId,
-          key: 'content',
-          type: 'updateMessage',
-          value: output,
-        });
+    // 如果是 function，则发送函数调用方法
+    if (isFunctionCall) {
+      const { function_call } = JSON.parse(output);
 
-        // 滚动到最后一条消息
-        const item = document.querySelector('#for-loading');
-        if (!item) return;
+      dispatchMessage({
+        id: assistantId,
+        key: 'function_call',
+        type: 'updateMessage',
+        value: function_call,
+      });
 
-        item.scrollIntoView({ behavior: 'smooth' });
-      },
-    });
+      await generateMessage(
+        [
+          ...messages,
+          { content: '', function_call, id: assistantId, role: 'assistant' } as ChatMessage,
+        ],
+        assistantId,
+        true,
+      );
+
+      dispatchMessage({
+        id: assistantId,
+        key: 'role',
+        type: 'updateMessage',
+        value: 'assistant',
+      });
+    }
   },
 
   resendMessage: async (messageId) => {
@@ -182,7 +235,7 @@ export const createChatSlice: StateCreator<
 
     const histories = chats
       .slice(0, currentIndex + 1)
-      // 如果点击重新发送的 message 其 role 是 assistant，那么需要移除
+      // 如果点击重新发送的 message 其 role 是 assistant 或者 function，那么需要移除
       // 如果点击重新发送的 message 其 role 是 user，则不需要移除
       .filter((c) => !(c.role === 'assistant' && c.id === messageId));
 

src/store/session/slices/chat/selectors/utils.ts

@@ -1,3 +1,4 @@
+import pluginList from '@/plugins';
 import { ChatMessage } from '@/types/chatMessage';
 import { LobeAgentSession } from '@/types/session';
 
@@ -5,6 +6,35 @@ export const organizeChats = (
   session: LobeAgentSession,
   avatar: { assistant: string; user: string },
 ) => {
+  const getMeta = (message: ChatMessage) => {
+    switch (message.role) {
+      case 'user': {
+        return {
+          avatar: avatar.user,
+        };
+      }
+
+      case 'system': {
+        return message.meta;
+      }
+
+      case 'assistant': {
+        return {
+          avatar: avatar.assistant,
+          title: session.meta.title,
+        };
+      }
+
+      case 'function': {
+        const plugin = pluginList.find((p) => p.name === message.function_call?.name);
+        return {
+          avatar: plugin?.avatar || '🧩',
+          title: plugin?.name || 'plugin-unknown',
+        };
+      }
+    }
+  };
+
   const basic = Object.values<ChatMessage>(session.chats)
     // 首先按照时间顺序排序，越早的在越前面
     .sort((pre, next) => pre.createAt - next.createAt)
@@ -14,17 +44,7 @@ export const organizeChats = (
     .map((m) => {
       return {
         ...m,
-        meta:
-          m.role === 'assistant'
-            ? {
-                avatar: avatar.assistant,
-                title: session.meta.title,
-              }
-            : m.role === 'user'
-            ? {
-                avatar: avatar.user,
-              }
-            : m.meta,
+        meta: getMeta(m),
       };
     });
 

src/types/chatMessage.ts

@@ -22,7 +22,6 @@ export interface ChatMessage extends BaseDataModel {
    */
   content: string;
   error?: any;
-
   // 扩展字段
   extra?: {
     fromModel?: string;
@@ -33,6 +32,9 @@ export interface ChatMessage extends BaseDataModel {
     };
   } & Record<string, any>;
 
+  function_call?: { arguments?: string; name: string };
+  name?: string;
+
   parentId?: string;
   // 引用
   quotaId?: string;

src/types/llm.ts

@@ -39,7 +39,7 @@ export interface LLMParams {
   top_p?: number;
 }
 
-export type LLMRoleType = 'user' | 'system' | 'assistant';
+export type LLMRoleType = 'user' | 'system' | 'assistant' | 'function';
 
 export interface LLMMessage {
   content: string;

src/utils/fetch.ts

@@ -91,7 +91,7 @@ interface FetchAITaskResultParams<T> {
 }
 
 export const fetchAIFactory =
-  <T>(fetcher: (params: T, signal?: AbortSignal) => Promise<Response>) =>
+  <T>(fetcher: (params: T, options: { signal?: AbortSignal }) => Promise<Response>) =>
   async ({
     params,
     onMessageHandle,
@@ -117,7 +117,7 @@ export const fetchAIFactory =
 
     onLoadingChange?.(true);
 
-    const data = await fetchSSE(() => fetcher(params, abortController?.signal), {
+    const data = await fetchSSE(() => fetcher(params, { signal: abortController?.signal }), {
       onErrorHandle: (error) => {
         errorHandle(new Error(error.message));
       },

src/utils/message.ts

@@ -0,0 +1,5 @@
+import { FUNCTION_MESSAGE_FLAG } from '@/const/message';
+
+export const isFunctionMessage = (content: string) => {
+  return content.startsWith(FUNCTION_MESSAGE_FLAG);
+};