feat: 知识库匹配模式选择

.dockerignore

@@ -8,3 +8,4 @@ README.md
 
 .yalc/
 yalc.lock
+testApi/

src/constants/model.ts

@@ -4,14 +4,12 @@ import type { RedisModelDataItemType } from '@/types/redis';
 export enum ChatModelNameEnum {
   GPT35 = 'gpt-3.5-turbo',
   VECTOR_GPT = 'VECTOR_GPT',
-  GPT3 = 'text-davinci-003',
   VECTOR = 'text-embedding-ada-002'
 }
 
 export const ChatModelNameMap = {
   [ChatModelNameEnum.GPT35]: 'gpt-3.5-turbo',
   [ChatModelNameEnum.VECTOR_GPT]: 'gpt-3.5-turbo',
-  [ChatModelNameEnum.GPT3]: 'text-davinci-003',
   [ChatModelNameEnum.VECTOR]: 'text-embedding-ada-002'
 };
 
@@ -34,7 +32,7 @@ export const modelList: ModelConstantsData[] = [
     trainName: '',
     maxToken: 4000,
     contextMaxToken: 7500,
-    maxTemperature: 2,
+    maxTemperature: 1.5,
     price: 3
   },
   {
@@ -47,16 +45,6 @@ export const modelList: ModelConstantsData[] = [
     maxTemperature: 1,
     price: 3
   }
-  // {
-  //   serviceCompany: 'openai',
-  //   name: 'GPT3',
-  //   model: ChatModelNameEnum.GPT3,
-  //   trainName: 'davinci',
-  //   maxToken: 4000,
-  //   contextMaxToken: 7500,
-  //   maxTemperature: 2,
-  //   price: 30
-  // }
 ];
 
 export enum TrainingStatusEnum {
@@ -97,6 +85,34 @@ export const ModelDataStatusMap: Record<RedisModelDataItemType['status'], string
   waiting: '训练中'
 };
 
+/* 知识库搜索时的配置 */
+// 搜索方式
+export enum ModelVectorSearchModeEnum {
+  hightSimilarity = 'hightSimilarity', // 高相似度+禁止回复
+  lowSimilarity = 'lowSimilarity', // 低相似度
+  noContext = 'noContex' // 高相似度+无上下文回复
+}
+export const ModelVectorSearchModeMap: Record<
+  `${ModelVectorSearchModeEnum}`,
+  {
+    text: string;
+    similarity: number;
+  }
+> = {
+  [ModelVectorSearchModeEnum.hightSimilarity]: {
+    text: '高相似度, 无匹配时拒绝回复',
+    similarity: 0.2
+  },
+  [ModelVectorSearchModeEnum.noContext]: {
+    text: '高相似度，无匹配时直接回复',
+    similarity: 0.2
+  },
+  [ModelVectorSearchModeEnum.lowSimilarity]: {
+    text: '低相似度匹配',
+    similarity: 0.8
+  }
+};
+
 export const defaultModel: ModelSchema = {
   _id: '',
   userId: '',
@@ -108,6 +124,9 @@ export const defaultModel: ModelSchema = {
   systemPrompt: '',
   intro: '',
   temperature: 5,
+  search: {
+    mode: ModelVectorSearchModeEnum.hightSimilarity
+  },
   service: {
     company: 'openai',
     trainId: '',

src/pages/api/chat/vectorGpt.ts

@@ -7,7 +7,7 @@ import { ChatItemType } from '@/types/chat';
 import { jsonRes } from '@/service/response';
 import type { ModelSchema } from '@/types/mongoSchema';
 import { PassThrough } from 'stream';
-import { modelList } from '@/constants/model';
+import { modelList, ModelVectorSearchModeMap, ModelVectorSearchModeEnum } from '@/constants/model';
 import { pushChatBill } from '@/service/events/pushBill';
 import { connectRedis } from '@/service/redis';
 import { VecModelDataPrefix } from '@/constants/redis';
@@ -65,13 +65,14 @@ export default async function handler(req: NextApiRequest, res: NextApiResponse)
       text: prompt.value
     });
 
+    const similarity = ModelVectorSearchModeMap[model.search.mode]?.similarity || 0.22;
     // 搜索系统提示词, 按相似度从 redis 中搜出相关的 q 和 text
     const redisData: any[] = await redis.sendCommand([
       'FT.SEARCH',
       `idx:${VecModelDataPrefix}:hash`,
       `@modelId:{${String(
         chat.modelId._id
-      )}} @vector:[VECTOR_RANGE 0.22 $blob]=>{$YIELD_DISTANCE_AS: score}`,
+      )}} @vector:[VECTOR_RANGE ${similarity} $blob]=>{$YIELD_DISTANCE_AS: score}`,
       'RETURN',
       '1',
       'text',
@@ -97,7 +98,24 @@ export default async function handler(req: NextApiRequest, res: NextApiResponse)
       }
     }
 
-    if (formatRedisPrompt.length > 0) {
+    /* 高相似度+退出，无法匹配时直接退出 */
+    if (
+      formatRedisPrompt.length === 0 &&
+      model.search.mode === ModelVectorSearchModeEnum.hightSimilarity
+    ) {
+      return res.send('对不起，你的问题不在知识库中。');
+    }
+    /* 高相似度+无上下文，不添加额外知识 */
+    if (
+      formatRedisPrompt.length === 0 &&
+      model.search.mode === ModelVectorSearchModeEnum.noContext
+    ) {
+      prompts.unshift({
+        obj: 'SYSTEM',
+        value: model.systemPrompt
+      });
+    } else {
+      // 有匹配情况下，添加知识库内容。
       // 系统提示词过滤，最多 2800 tokens
       const systemPrompt = systemPromptFilter(formatRedisPrompt, 2800);
 
@@ -107,8 +125,6 @@ export default async function handler(req: NextApiRequest, res: NextApiResponse)
           'YYYY/MM/DD HH:mm:ss'
         )} ${systemPrompt}"`
       });
-    } else {
-      return res.send('对不起，你的问题不在知识库中。');
     }
 
     // 控制在 tokens 数量，防止超出

src/pages/api/model/update.ts

@@ -8,7 +8,7 @@ import type { ModelUpdateParams } from '@/types/model';
 /* 获取我的模型 */
 export default async function handler(req: NextApiRequest, res: NextApiResponse<any>) {
   try {
-    const { name, service, security, systemPrompt, intro, temperature } =
+    const { name, search, service, security, systemPrompt, intro, temperature } =
       req.body as ModelUpdateParams;
     const { modelId } = req.query as { modelId: string };
     const { authorization } = req.headers;
@@ -37,6 +37,7 @@ export default async function handler(req: NextApiRequest, res: NextApiResponse<
         systemPrompt,
         intro,
         temperature,
+        search,
         // service,
         security
       }

src/pages/api/openapi/chat/lafGpt.ts

@@ -83,22 +83,22 @@ export default async function handler(req: NextApiRequest, res: NextApiResponse)
 下面是一些例子:
 实现一个手机号发生注册验证码方法.
 1. 从 query 中获取 phone.
-2. 校验手机号格式是否正确,不正确返回{error: "手机号格式错误"}.
+2. 校验手机号格式是否正确,不正确则返回错误响应,消息为:手机号格式错误.
 3. 给 phone 发送一个短信验证码,验证码长度为6位字符串,内容为:你正在注册laf,验证码为:code.
 4. 数据库添加数据,表为"codes",内容为 {phone, code}.
 
 实现根据手机号注册账号,需要验证手机验证码.
 1. 从 body 中获取 phone 和 code.
-2. 校验手机号格式是否正确,不正确返回{error: "手机号格式错误"}.
+2. 校验手机号格式是否正确,不正确返回错误响应,消息为:手机号格式错误.
-2. 获取数据库数据,表为"codes",查找是否有符合 phone, code 等于body参数的记录,没有的话返回 {error:"验证码不正确"}.
+2. 获取数据库数据,表为"codes",查找是否有符合 phone, code 等于body参数的记录,没有的话错误响应,消息为:验证码不正确.
 4. 添加数据库数据,表为"users" ,内容为{phone, code, createTime}.
 5. 删除数据库数据,删除 code 记录.
 
 更新博客记录。传入blogId,blogText,tags,还需要记录更新的时间.
 1. 从 body 中获取 blogId,blogText 和 tags.
-2. 校验 blogId 是否为空,为空则返回 {error: "博客ID不能为空"}.
+2. 校验 blogId 是否为空,为空则错误响应,消息为:博客ID不能为空.
-3. 校验 blogText 是否为空,为空则返回 {error: "博客内容不能为空"}.
+3. 校验 blogText 是否为空,为空则错误响应,消息为:博客内容不能为空.
-4. 校验 tags 是否为数组,不是则返回 {error: "标签必须为数组"}.
+4. 校验 tags 是否为数组,不是则错误响应,消息为:标签必须为数组.
 5. 获取当前时间,记录为 updateTime.
 6. 更新数据库数据,表为"blogs",更新符合 blogId 的记录的内容为{blogText, tags, updateTime}.
 7. 返回结果 {message: "更新博客记录成功"}.`

src/pages/chat/index.tsx

@@ -114,8 +114,7 @@ const Chat = ({ chatId }: { chatId: string }) => {
     async (prompts: ChatSiteItemType) => {
       const urlMap: Record<string, string> = {
         [ChatModelNameEnum.GPT35]: '/api/chat/chatGpt',
-        [ChatModelNameEnum.VECTOR_GPT]: '/api/chat/vectorGpt',
+        [ChatModelNameEnum.VECTOR_GPT]: '/api/chat/vectorGpt'
-        [ChatModelNameEnum.GPT3]: '/api/chat/gpt3'
       };
 
       if (!urlMap[chatData.modelName]) return Promise.reject('找不到模型');

src/pages/model/detail/components/ModelEditForm.tsx

@@ -12,12 +12,13 @@ import {
   SliderThumb,
   SliderMark,
   Tooltip,
-  Button
+  Button,
+  Select
 } from '@chakra-ui/react';
 import { QuestionOutlineIcon } from '@chakra-ui/icons';
 import type { ModelSchema } from '@/types/mongoSchema';
 import { UseFormReturn } from 'react-hook-form';
-import { modelList } from '@/constants/model';
+import { modelList, ModelVectorSearchModeMap } from '@/constants/model';
 import { formatPrice } from '@/utils/user';
 import { useConfirm } from '@/hooks/useConfirm';
 
@@ -89,15 +90,6 @@ const ModelEditForm = ({
             删除模型
           </Button>
         </Flex>
-        {/* <FormControl mt={4}>
-          <Box mb={1}>介绍:</Box>
-          <Textarea
-            rows={5}
-            maxLength={500}
-            {...register('intro')}
-            placeholder={'模型的介绍，仅做展示，不影响模型的效果'}
-          />
-        </FormControl> */}
       </Card>
       <Card p={4}>
         <Box fontWeight={'bold'}>模型效果</Box>
@@ -143,6 +135,20 @@ const ModelEditForm = ({
             </Slider>
           </Flex>
         </FormControl>
+        {canTrain && (
+          <FormControl mt={4}>
+            <Flex alignItems={'center'}>
+              <Box flex={'0 0 70px'}>搜索模式</Box>
+              <Select {...register('search.mode', { required: '搜索模式不能为空' })}>
+                {Object.entries(ModelVectorSearchModeMap).map(([key, { text }]) => (
+                  <option key={key} value={key}>
+                    {text}
+                  </option>
+                ))}
+              </Select>
+            </Flex>
+          </FormControl>
+        )}
         <Box mt={4}>
           <Box mb={1}>系统提示词</Box>
           <Textarea

src/pages/model/detail/index.tsx

@@ -143,6 +143,7 @@ const ModelDetail = ({ modelId }: { modelId: string }) => {
           systemPrompt: data.systemPrompt,
           intro: data.intro,
           temperature: data.temperature,
+          search: data.search,
           service: data.service,
           security: data.security
         });

src/service/models/model.ts

@@ -1,5 +1,7 @@
 import { Schema, model, models, Model as MongoModel } from 'mongoose';
 import { ModelSchema as ModelType } from '@/types/mongoSchema';
+import { ModelVectorSearchModeMap, ModelVectorSearchModeEnum } from '@/constants/model';
+
 const ModelSchema = new Schema({
   userId: {
     type: Schema.Types.ObjectId,
@@ -43,6 +45,13 @@ const ModelSchema = new Schema({
     max: 10,
     default: 4
   },
+  search: {
+    mode: {
+      type: String,
+      enum: Object.keys(ModelVectorSearchModeMap),
+      default: ModelVectorSearchModeEnum.hightSimilarity
+    }
+  },
   service: {
     company: {
       type: String,

src/types/model.d.ts

@@ -5,8 +5,9 @@ export interface ModelUpdateParams {
   systemPrompt: string;
   intro: string;
   temperature: number;
-  service: ModelSchema.service;
+  search: ModelSchema['search'];
-  security: ModelSchema.security;
+  service: ModelSchema['service'];
+  security: ModelSchema['security'];
 }
 
 export interface ModelDataItemType {

src/types/mongoSchema.d.ts

@@ -1,5 +1,10 @@
 import type { ChatItemType } from './chat';
-import { ModelStatusEnum, TrainingStatusEnum, ChatModelNameEnum } from '@/constants/model';
+import {
+  ModelStatusEnum,
+  TrainingStatusEnum,
+  ChatModelNameEnum,
+  ModelVectorSearchModeEnum
+} from '@/constants/model';
 import type { DataType } from './data';
 
 export type ServiceName = 'openai';
@@ -32,6 +37,9 @@ export interface ModelSchema {
   updateTime: number;
   trainingTimes: number;
   temperature: number;
+  search: {
+    mode: `${ModelVectorSearchModeEnum}`;
+  };
   service: {
     company: ServiceName;
     trainId: string; // 训练的模型，训练后就是训练的模型id