perf: kb-add last question to search

public/docs/versionIntro.md

@@ -1,5 +1,6 @@
 ### Fast GPT V3.1
 
+- 优化 - 知识库搜索，会将上一个问题并入搜索范围。
 - 优化 - 模型结构设计，不再区分知识库和对话模型，而是通过开关的形式，手动选择手否需要进行知识库搜索。
 - 新增 - 模型共享市场，可以使用其他用户分享的模型。
 - 新增 - 邀请好友注册功能。

src/pages/api/chat/chat.ts

@@ -58,7 +58,7 @@ export default async function handler(req: NextApiRequest, res: NextApiResponse)
       const { code, searchPrompt } = await searchKb({
         userApiKey,
         systemApiKey,
-        text: prompt.value,
+        prompts,
         similarity: ModelVectorSearchModeMap[model.chat.searchMode]?.similarity,
         model,
         userId

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

@@ -66,7 +66,7 @@ export default async function handler(req: NextApiRequest, res: NextApiResponse)
 
       const { code, searchPrompt } = await searchKb({
         systemApiKey: apiKey,
-        text: prompts[prompts.length - 1].value,
+        prompts,
         similarity,
         model,
         userId

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

@@ -118,7 +118,7 @@ export default async function handler(req: NextApiRequest, res: NextApiResponse)
     const { searchPrompt } = await searchKb({
       systemApiKey: apiKey,
       similarity: ModelVectorSearchModeMap[model.chat.searchMode]?.similarity,
-      text: prompt.value,
+      prompts,
       model,
       userId
     });

src/service/events/generateVector.ts

@@ -60,8 +60,8 @@ export async function generateVector(next = false): Promise<any> {
     }
 
     // 生成词向量
-    const { vector } = await openaiCreateEmbedding({
+    const { vectors } = await openaiCreateEmbedding({
-      text: dataItem.q,
+      textArr: [dataItem.q],
       userId: dataItem.userId,
       userApiKey,
       systemApiKey
@@ -70,7 +70,7 @@ export async function generateVector(next = false): Promise<any> {
     // 更新 pg 向量和状态数据
     await PgClient.update('modelData', {
       values: [
-        { key: 'vector', value: `[${vector}]` },
+        { key: 'vector', value: `[${vectors[0]}]` },
         { key: 'status', value: `ready` }
       ],
       where: [['id', dataId]]

src/service/plugins/searchKb.ts

@@ -4,6 +4,7 @@ import { ModelSchema } from '@/types/mongoSchema';
 import { openaiCreateEmbedding } from '../utils/chat/openai';
 import { ChatRoleEnum } from '@/constants/chat';
 import { sliceTextByToken } from '@/utils/chat';
+import { ChatItemSimpleType } from '@/types/chat';
 
 /**
  *  use openai embedding search kb
@@ -11,14 +12,14 @@ import { sliceTextByToken } from '@/utils/chat';
 export const searchKb = async ({
   userApiKey,
   systemApiKey,
-  text,
+  prompts,
   similarity = 0.2,
   model,
   userId
 }: {
   userApiKey?: string;
   systemApiKey: string;
-  text: string;
+  prompts: ChatItemSimpleType[];
   model: ModelSchema;
   userId: string;
   similarity?: number;
@@ -29,18 +30,19 @@ export const searchKb = async ({
     value: string;
   };
 }> => {
-  const modelConstantsData = ChatModelMap[model.chat.chatModel];
+  async function search(textArr: string[] = []) {
-
     // 获取提示词的向量
-  const { vector: promptVector } = await openaiCreateEmbedding({
+    const { vectors: promptVectors } = await openaiCreateEmbedding({
       userApiKey,
       systemApiKey,
       userId,
-    text
+      textArr
     });
 
-  const vectorSearch = await PgClient.select<{ q: string; a: string }>('modelData', {
+    const searchRes = await Promise.all(
-    fields: ['q', 'a'],
+      promptVectors.map((promptVector) =>
+        PgClient.select<{ id: string; q: string; a: string }>('modelData', {
+          fields: ['id', 'q', 'a'],
           where: [
             ['status', ModelDataStatusEnum.ready],
             'AND',
@@ -50,9 +52,34 @@ export const searchKb = async ({
           ],
           order: [{ field: 'vector', mode: `<=> '[${promptVector}]'` }],
           limit: 20
-  });
+        }).then((res) => res.rows)
+      )
+    );
 
-  const systemPrompts: string[] = vectorSearch.rows.map((item) => `${item.q}\n${item.a}`);
+    // Remove repeat record
+    const idSet = new Set<string>();
+    const filterSearch = searchRes.map((search) =>
+      search.filter((item) => {
+        if (idSet.has(item.id)) {
+          return false;
+        }
+        idSet.add(item.id);
+        return true;
+      })
+    );
+
+    return filterSearch.map((item) => item.map((item) => `${item.q}\n${item.a}`).join('\n'));
+  }
+  const modelConstantsData = ChatModelMap[model.chat.chatModel];
+
+  // search three times
+  const userPrompts = prompts.filter((item) => item.obj === 'Human');
+
+  const searchArr: string[] = [
+    userPrompts[userPrompts.length - 1].value,
+    userPrompts[userPrompts.length - 2]?.value
+  ].filter((item) => item);
+  const systemPrompts = await search(searchArr);
 
   // filter system prompt
   if (
@@ -80,13 +107,24 @@ export const searchKb = async ({
     };
   }
 
-  // 有匹配情况下，system 添加知识库内容。
+  /* 有匹配情况下，system 添加知识库内容。 */
-  // 系统提示词过滤，最多 65% tokens
+
-  const filterSystemPrompt = sliceTextByToken({
+  // filter system prompts. max 70% tokens
+  const filterRateMap: Record<number, number[]> = {
+    1: [0.7],
+    2: [0.5, 0.2]
+  };
+  const filterRate = filterRateMap[systemPrompts.length] || filterRateMap[0];
+
+  const filterSystemPrompt = filterRate
+    .map((rate, i) =>
+      sliceTextByToken({
         model: model.chat.chatModel,
-    text: systemPrompts.join('\n'),
+        text: systemPrompts[i],
-    length: Math.floor(modelConstantsData.contextMaxToken * 0.65)
+        length: Math.floor(modelConstantsData.contextMaxToken * rate)
-  });
+      })
+    )
+    .join('\n');
 
   return {
     code: 200,

src/service/utils/chat/openai.ts

@@ -22,12 +22,12 @@ export const openaiCreateEmbedding = async ({
   userApiKey,
   systemApiKey,
   userId,
-  text
+  textArr
 }: {
   userApiKey?: string;
   systemApiKey: string;
   userId: string;
-  text: string;
+  textArr: string[];
 }) => {
   // 获取 chatAPI
   const chatAPI = getOpenAIApi(userApiKey || systemApiKey);
@@ -37,7 +37,7 @@ export const openaiCreateEmbedding = async ({
     .createEmbedding(
       {
         model: embeddingModel,
-        input: text
+        input: textArr
       },
       {
         timeout: 60000,
@@ -46,18 +46,18 @@ export const openaiCreateEmbedding = async ({
     )
     .then((res) => ({
       tokenLen: res.data.usage.total_tokens || 0,
-      vector: res.data.data?.[0]?.embedding || []
+      vectors: res.data.data.map((item) => item.embedding)
     }));
 
   pushGenerateVectorBill({
     isPay: !userApiKey,
     userId,
-    text,
+    text: textArr.join(''),
     tokenLen: res.tokenLen
   });
 
   return {
-    vector: res.vector,
+    vectors: res.vectors,
     chatAPI
   };
 };