perf: token slice

README.md

@@ -232,6 +232,7 @@ services:
 #### 3. 运行 docker-compose
 
 下面是一个辅助脚本，也可以直接 docker-compose up -d
+
 **run.sh 运行文件**
 
 ```bash

src/api/fetch.ts

@@ -25,7 +25,7 @@ export const streamFetch = ({ url, data, onMessage, abortSignal }: StreamFetchPr
 
         const decoder = new TextDecoder();
 
-        const systemPrompt = decodeURIComponent(res.headers.get(SYSTEM_PROMPT_HEADER) || '');
+        const systemPrompt = decodeURIComponent(res.headers.get(SYSTEM_PROMPT_HEADER) || '').trim();
         const newChatId = decodeURIComponent(res.headers.get(NEW_CHATID_HEADER) || '');
 
         let responseText = '';

src/constants/model.ts

@@ -28,7 +28,7 @@ export const ChatModelMap = {
     chatModel: OpenAiChatEnum.GPT35,
     name: 'ChatGpt',
     contextMaxToken: 4096,
-    systemMaxToken: 3000,
+    systemMaxToken: 2500,
     maxTemperature: 1.5,
     price: 3
   },
@@ -36,7 +36,7 @@ export const ChatModelMap = {
     chatModel: OpenAiChatEnum.GPT4,
     name: 'Gpt4',
     contextMaxToken: 8000,
-    systemMaxToken: 4000,
+    systemMaxToken: 3500,
     maxTemperature: 1.5,
     price: 30
   },
@@ -44,7 +44,7 @@ export const ChatModelMap = {
     chatModel: OpenAiChatEnum.GPT432k,
     name: 'Gpt4-32k',
     contextMaxToken: 32000,
-    systemMaxToken: 4000,
+    systemMaxToken: 6000,
     maxTemperature: 1.5,
     price: 30
   },

src/pages/chat/index.tsx

@@ -643,7 +643,7 @@ const Chat = ({ modelId, chatId }: { modelId: string; chatId: string }) => {
           <ModalOverlay />
           <ModalContent maxW={'min(90vw, 600px)'} pr={2} maxH={'80vh'} overflowY={'auto'}>
             <ModalCloseButton />
-            <ModalBody pt={10} fontSize={'sm'} whiteSpace={'pre-wrap'} textAlign={'justify'}>
+            <ModalBody pt={5} fontSize={'sm'} whiteSpace={'pre-wrap'} textAlign={'justify'}>
               {showSystemPrompt}
             </ModalBody>
           </ModalContent>

src/service/plugins/searchKb.ts

@@ -85,11 +85,24 @@ export const searchKb = async ({
   };
   const filterRate = filterRateMap[systemPrompts.length] || filterRateMap[0];
 
+  // count fixed system prompt
+  const fixedSystemPrompt = `
+${model.chat.systemPrompt}
+${
+  model.chat.searchMode === ModelVectorSearchModeEnum.hightSimilarity ? '不回答知识库外的内容.' : ''
+}
+知识库内容为:`;
+  const fixedSystemTokens = modelToolMap[model.chat.chatModel].countTokens({
+    messages: [{ obj: 'System', value: fixedSystemPrompt }]
+  });
+
+  const maxTokens = modelConstantsData.systemMaxToken - fixedSystemTokens;
+
   const filterSystemPrompt = filterRate
     .map((rate, i) =>
       modelToolMap[model.chat.chatModel].sliceText({
         text: systemPrompts[i],
-        length: Math.floor(modelConstantsData.systemMaxToken * rate)
+        length: Math.floor(maxTokens * rate)
       })
     )
     .join('\n');
@@ -122,13 +135,7 @@ export const searchKb = async ({
     code: 200,
     searchPrompt: {
       obj: ChatRoleEnum.System,
-      value: `
-${model.chat.systemPrompt}
-${
-  model.chat.searchMode === ModelVectorSearchModeEnum.hightSimilarity ? '不回答知识库外的内容.' : ''
-}
-知识库内容为: '${filterSystemPrompt}'
-`
+      value: `${fixedSystemPrompt}'${filterSystemPrompt}'`
     }
   };
 };

src/service/utils/chat/claude.ts

@@ -45,18 +45,13 @@ export const lafClaudChat = async ({
     }
   );
 
-  let responseText = '';
-  let totalTokens = 0;
-
-  if (!stream) {
-    responseText = lafResponse.data?.text || '';
-  }
+  const responseText = stream ? '' : lafResponse.data?.text || '';
 
   return {
     streamResponse: lafResponse,
     responseMessages: messages.concat({ obj: ChatRoleEnum.AI, value: responseText }),
     responseText,
-    totalTokens
+    totalTokens: 0
   };
 };
 
@@ -83,18 +78,15 @@ export const lafClaudStreamResponse = async ({
     } catch (error) {
       console.log('pipe error', error);
     }
-    // count tokens
+
     const finishMessages = prompts.concat({
       obj: ChatRoleEnum.AI,
       value: responseContent
     });
-    const totalTokens = modelToolMap[ClaudeEnum.Claude].countTokens({
-      messages: finishMessages
-    });
 
     return {
       responseContent,
-      totalTokens,
+      totalTokens: 0,
       finishMessages
     };
   } catch (error) {

src/service/utils/chat/openai.ts

@@ -96,14 +96,8 @@ export const chatResponse = async ({
     }
   );
 
-  let responseText = '';
-  let totalTokens = 0;
-
-  // adapt data
-  if (!stream) {
-    responseText = response.data.choices[0].message?.content || '';
-    totalTokens = response.data.usage?.total_tokens || 0;
-  }
+  const responseText = stream ? '' : response.data.choices[0].message?.content || '';
+  const totalTokens = stream ? 0 : response.data.usage?.total_tokens || 0;
 
   return {
     streamResponse: response,

src/utils/chat/index.ts

@@ -17,14 +17,14 @@ export const modelToolMap: Record<
   },
   [OpenAiChatEnum.GPT4]: {
     countTokens: ({ messages }) => countOpenAIToken({ model: OpenAiChatEnum.GPT4, messages }),
-    sliceText: (data) => openAiSliceTextByToken({ model: OpenAiChatEnum.GPT35, ...data })
+    sliceText: (data) => openAiSliceTextByToken({ model: OpenAiChatEnum.GPT4, ...data })
   },
   [OpenAiChatEnum.GPT432k]: {
     countTokens: ({ messages }) => countOpenAIToken({ model: OpenAiChatEnum.GPT432k, messages }),
-    sliceText: (data) => openAiSliceTextByToken({ model: OpenAiChatEnum.GPT35, ...data })
+    sliceText: (data) => openAiSliceTextByToken({ model: OpenAiChatEnum.GPT432k, ...data })
   },
   [ClaudeEnum.Claude]: {
-    countTokens: () => 0,
-    sliceText: ClaudeSliceTextByToken
+    countTokens: ({ messages }) => countOpenAIToken({ model: OpenAiChatEnum.GPT35, messages }),
+    sliceText: (data) => openAiSliceTextByToken({ model: OpenAiChatEnum.GPT35, ...data })
   }
 };