feat: 摘要拆分

src/service/events/generateAbstract.ts

@@ -0,0 +1,177 @@
+import { DataItem } from '@/service/mongo';
+import { getOpenAIApi } from '@/service/utils/chat';
+import { httpsAgent, getOpenApiKey } from '@/service/utils/tools';
+import type { ChatCompletionRequestMessage } from 'openai';
+import { DataItemSchema } from '@/types/mongoSchema';
+import { ChatModelNameEnum } from '@/constants/model';
+import { pushSplitDataBill } from '@/service/events/pushBill';
+
+export async function generateAbstract(next = false): Promise<any> {
+  if (global.generatingAbstract && !next) return;
+  global.generatingAbstract = true;
+
+  const systemPrompt: ChatCompletionRequestMessage = {
+    role: 'system',
+    content: `我会向你发送一段长文本，请从中总结出3~10个摘要，尽量详细，请按以下格式返回: "(1):"\n"(2):"\n"(3):"\n`
+  };
+  let dataItem: DataItemSchema | null = null;
+
+  try {
+    // 找出一个需要生成的 dataItem
+    dataItem = await DataItem.findOne({
+      status: { $ne: 0 },
+      times: { $gt: 0 },
+      type: 'abstract'
+    });
+
+    if (!dataItem) {
+      console.log('没有需要生成 【摘要】 的数据');
+      global.generatingAbstract = false;
+      return;
+    }
+
+    // 更新状态为生成中
+    await DataItem.findByIdAndUpdate(dataItem._id, {
+      status: 2
+    });
+
+    // 获取 openapi Key
+    let userApiKey, systemKey;
+    try {
+      const key = await getOpenApiKey(dataItem.userId);
+      userApiKey = key.userApiKey;
+      systemKey = key.systemKey;
+    } catch (error) {
+      // 余额不够了, 把用户所有记录改成闲置
+      await DataItem.updateMany({
+        userId: dataItem.userId,
+        status: 0
+      });
+      throw new Error('获取 openai key 失败');
+    }
+
+    console.log('正在生成一组摘要, ID:', dataItem._id);
+
+    const startTime = Date.now();
+
+    // 获取 openai 请求实例
+    const chatAPI = getOpenAIApi(userApiKey || systemKey);
+    // 请求 chatgpt 获取摘要
+    const abstractResponse = await Promise.allSettled(
+      [0.5, 1].map((temperature) =>
+        chatAPI.createChatCompletion(
+          {
+            model: ChatModelNameEnum.GPT35,
+            temperature: temperature,
+            n: 1,
+            messages: [
+              systemPrompt,
+              {
+                role: 'user',
+                content: dataItem?.text || ''
+              }
+            ]
+          },
+          {
+            timeout: 120000,
+            httpsAgent
+          }
+        )
+      )
+    );
+
+    // 过滤出成功的响应
+    const successAbstracts = abstractResponse.filter((item) => item.status === 'fulfilled');
+    // 提取摘要内容
+    const rawContents: string[] = successAbstracts.map(
+      (item: any) => item?.value?.data.choices[0].message?.content || ''
+    );
+    // 从 content 中提取摘要内容
+    const splitContents = rawContents.map((content) => splitText(content)).flat();
+
+    // 生成词向量
+    const vectorResponse = await Promise.allSettled(
+      splitContents.map((item) =>
+        chatAPI.createEmbedding({
+          model: 'text-embedding-ada-002',
+          input: item.abstract
+        })
+      )
+    );
+    // 筛选成功的向量请求
+    const vectorSuccessResponse = vectorResponse
+      .map((item: any, i) => {
+        if (item.status !== 'fulfilled') return '';
+        return {
+          abstract: splitContents[i].abstract,
+          abstractVector: item?.value?.data?.data?.[0]?.embedding
+        };
+      })
+      .filter((item) => item);
+
+    // 插入数据库，并修改状态
+    await DataItem.findByIdAndUpdate(dataItem._id, {
+      status: 0,
+      $push: {
+        rawResponse: {
+          $each: rawContents
+        },
+        result: {
+          $each: vectorSuccessResponse
+        }
+      }
+    });
+
+    // 计费
+    !userApiKey &&
+      splitContents.length > 0 &&
+      pushSplitDataBill({
+        userId: dataItem.userId,
+        type: 'abstract',
+        text:
+          systemPrompt.content +
+          dataItem.text +
+          rawContents.join('') +
+          rawContents.join('').substring(0, Math.floor(dataItem.text.length / 10)) // 向量价格是gpt35的1/10
+      });
+    console.log(
+      '生成摘要成功，time:',
+      `${(Date.now() - startTime) / 1000}s`,
+      '摘要数量：',
+      splitContents.length
+    );
+  } catch (error: any) {
+    console.log('error: 生成摘要错误', dataItem?._id);
+    console.log('response:', error);
+    if (dataItem?._id) {
+      await DataItem.findByIdAndUpdate(dataItem._id, {
+        status: dataItem.times > 0 ? 1 : 0, // 还有重试次数则可以继续进行
+        $inc: {
+          // 剩余尝试次数-1
+          times: -1
+        }
+      });
+    }
+  }
+
+  generateAbstract(true);
+}
+
+/**
+ * 检查文本是否按格式返回
+ */
+function splitText(text: string) {
+  const regex = /\(\d+\):(\s*)(.*)(\s*)/g;
+  const matches = text.matchAll(regex); // 获取所有匹配到的结果
+
+  const result = []; // 存储最终的结果
+  for (const match of matches) {
+    if (match[2]) {
+      result.push({
+        abstract: match[2] as string
+      });
+    }
+  }
+
+  return result;
+}

src/service/events/generateQA.ts

@@ -20,7 +20,8 @@ export async function generateQA(next = false): Promise<any> {
     // 找出一个需要生成的 dataItem
     dataItem = await DataItem.findOne({
       status: { $ne: 0 },
-      times: { $gt: 0 }
+      times: { $gt: 0 },
+      type: 'QA'
     });
 
     if (!dataItem) {
@@ -49,62 +50,72 @@ export async function generateQA(next = false): Promise<any> {
       throw new Error('获取 openai key 失败');
     }
 
-    console.log('正在生成一个QA, ID:', dataItem._id, 'temperature: ', dataItem.temperature / 100);
+    console.log('正在生成一组QA, ID:', dataItem._id);
 
     const startTime = Date.now();
 
     // 获取 openai 请求实例
     const chatAPI = getOpenAIApi(userApiKey || systemKey);
     // 请求 chatgpt 获取回答
-    const response = await chatAPI.createChatCompletion(
-      {
-        model: ChatModelNameEnum.GPT35,
-        temperature: dataItem.temperature / 100,
-        n: 1,
-        messages: [
-          systemPrompt,
+    const response = await Promise.allSettled(
+      [0, 0.5, 0.8].map((temperature) =>
+        chatAPI.createChatCompletion(
           {
-            role: 'user',
-            content: dataItem.text
+            model: ChatModelNameEnum.GPT35,
+            temperature: temperature,
+            n: 1,
+            messages: [
+              systemPrompt,
+              {
+                role: 'user',
+                content: dataItem?.text || ''
+              }
+            ]
+          },
+          {
+            timeout: 120000,
+            httpsAgent
           }
-        ]
-      },
-      {
-        timeout: 120000,
-        httpsAgent
-      }
+        )
+      )
+    );
+    // 过滤出成功的响应
+    const successResponse = response.filter((item) => item.status === 'fulfilled');
+    // 提取响应内容
+    const rawContents: string[] = successResponse.map(
+      (item: any) => item?.value?.data.choices[0].message?.content || ''
     );
-    const content = response.data.choices[0].message?.content;
     // 从 content 中提取 QA
-    const splitResponse = splitText(content || '');
+    const splitResponses = rawContents.map((content) => splitText(content)).flat();
     // 插入数据库，并修改状态
     await DataItem.findByIdAndUpdate(dataItem._id, {
-      status: dataItem.temperature >= 90 ? 0 : 1, // 需要生成 4 组内容。0,0.3,0.6,0.9
-      temperature: dataItem.temperature >= 90 ? dataItem.temperature : dataItem.temperature + 30,
+      status: 0,
       $push: {
-        rawResponse: content,
+        rawResponse: {
+          $each: rawContents
+        },
         result: {
-          $each: splitResponse
+          $each: splitResponses
         }
       }
     });
     // 计费
     !userApiKey &&
-      splitResponse.length > 0 &&
+      splitResponses.length > 0 &&
       pushSplitDataBill({
         userId: dataItem.userId,
-        text: systemPrompt.content + dataItem.text + content
+        type: 'QA',
+        text: systemPrompt.content + dataItem.text + rawContents.join('')
       });
     console.log(
       '生成QA成功，time:',
       `${(Date.now() - startTime) / 1000}s`,
       'QA数量：',
-      splitResponse.length
+      splitResponses.length
     );
   } catch (error: any) {
     console.log('error: 生成QA错误', dataItem?._id);
     console.log('response:', error?.response);
-    // 重置状态
     if (dataItem?._id) {
       await DataItem.findByIdAndUpdate(dataItem._id, {
         status: dataItem.times > 0 ? 1 : 0, // 还有重试次数则可以继续进行

src/service/events/pushBill.ts

@@ -2,6 +2,7 @@ import { connectToDatabase, Bill, User } from '../mongo';
 import { modelList, ChatModelNameEnum } from '@/constants/model';
 import { encode } from 'gpt-token-utils';
 import { formatPrice } from '@/utils/user';
+import type { DataType } from '@/types/data';
 
 export const pushChatBill = async ({
   modelName,
@@ -59,7 +60,15 @@ export const pushChatBill = async ({
   }
 };
 
-export const pushSplitDataBill = async ({ userId, text }: { userId: string; text: string }) => {
+export const pushSplitDataBill = async ({
+  userId,
+  text,
+  type
+}: {
+  userId: string;
+  text: string;
+  type: DataType;
+}) => {
   await connectToDatabase();
 
   let billId;
@@ -83,7 +92,7 @@ export const pushSplitDataBill = async ({ userId, text }: { userId: string; text
       // 插入 Bill 记录
       const res = await Bill.create({
         userId,
-        type: 'splitData',
+        type,
         modelName: ChatModelNameEnum.GPT35,
         textLen: text.length,
         tokenLen: tokens.length,

src/service/models/data.ts

@@ -1,5 +1,6 @@
 import { Schema, model, models, Model } from 'mongoose';
-import { DataItemSchema as Datatype } from '@/types/mongoSchema';
+import { DataSchema as Datatype } from '@/types/mongoSchema';
+import { DataTypeTextMap } from '@/constants/data';
 
 const DataSchema = new Schema({
   userId: {
@@ -15,6 +16,11 @@ const DataSchema = new Schema({
     type: Date,
     default: () => new Date()
   },
+  type: {
+    type: String,
+    required: true,
+    enum: Object.keys(DataTypeTextMap)
+  },
   isDeleted: {
     type: Boolean,
     default: false

src/service/models/dataItem.ts

@@ -1,5 +1,6 @@
 import type { DataItemSchema as DataItemType } from '@/types/mongoSchema';
 import { Schema, model, models, Model } from 'mongoose';
+import { DataTypeTextMap } from '@/constants/data';
 
 const DataItemSchema = new Schema({
   userId: {
@@ -12,19 +13,23 @@ const DataItemSchema = new Schema({
     ref: 'data',
     required: true
   },
+  type: {
+    type: String,
+    required: true,
+    enum: Object.keys(DataTypeTextMap)
+  },
   times: {
+    // 剩余重试次数
     type: Number,
     default: 3
   },
   text: {
+    // 文本内容
     type: String,
     required: true
   },
-  temperature: {
-    type: Number,
-    required: true
-  },
   rawResponse: {
+    // 原始拆分结果
     type: [String],
     default: []
   },
@@ -33,11 +38,21 @@ const DataItemSchema = new Schema({
       {
         q: {
           type: String,
-          required: true
+          default: ''
         },
         a: {
           type: String,
-          required: true
+          default: ''
+        },
+        abstract: {
+          // 摘要
+          type: String,
+          default: ''
+        },
+        abstractVector: {
+          // 摘要对应的向量
+          type: [Number],
+          default: []
         }
       }
     ],

src/service/mongo.ts

@@ -1,5 +1,7 @@
 import mongoose from 'mongoose';
 import { generateQA } from './events/generateQA';
+import { generateAbstract } from './events/generateAbstract';
+
 /**
  * 连接 MongoDB 数据库
  */
@@ -24,8 +26,8 @@ export async function connectToDatabase(): Promise<void> {
     global.mongodb = null;
   }
 
-  // 递归 QA 生成
   generateQA();
+  generateAbstract();
 }
 
 export * from './models/authCode';