V4.8.17 feature (#3493)

* split tokens into input and output (#3477) * split tokens into input and output * query extension & tool call & question guide * fix * perf: input and output tokens * perf: tool call if else * perf: remove code * fix: extract usage count * fix: qa usage count --------- Co-authored-by: heheer <heheer@sealos.io>
2025-07-24 05:23:57 +00:00 · 2024-12-30 10:13:25 +08:00
parent da2831b948
commit 50bf7f9a3b
46 changed files with 467 additions and 230 deletions
--- a/packages/service/core/ai/functions/createQuestionGuide.ts
+++ b/packages/service/core/ai/functions/createQuestionGuide.ts
@@ -1,6 +1,6 @@
 import type { ChatCompletionMessageParam } from '@fastgpt/global/core/ai/type.d';
 import { createChatCompletion } from '../config';
-import { countGptMessagesTokens } from '../../../common/string/tiktoken/index';
+import { countGptMessagesTokens, countPromptTokens } from '../../../common/string/tiktoken/index';
 import { loadRequestMessages } from '../../chat/utils';
 import { llmCompletionsBodyFormat } from '../utils';
 import {
@@ -20,7 +20,8 @@ export async function createQuestionGuide({
  customPrompt?: string;
 }): Promise<{
  result: string[];
-  tokens: number;
+  inputTokens: number;
+  outputTokens: number;
 }> {
  const concatMessages: ChatCompletionMessageParam[] = [
    ...messages,
@@ -29,6 +30,10 @@ export async function createQuestionGuide({
      content: `${customPrompt || PROMPT_QUESTION_GUIDE}\n${PROMPT_QUESTION_GUIDE_FOOTER}`
    }
  ];
+  const requestMessages = await loadRequestMessages({
+    messages: concatMessages,
+    useVision: false
+  });

  const { response: data } = await createChatCompletion({
    body: llmCompletionsBodyFormat(
@@ -36,10 +41,7 @@ export async function createQuestionGuide({
        model,
        temperature: 0.1,
        max_tokens: 200,
-        messages: await loadRequestMessages({
-          messages: concatMessages,
-          useVision: false
-        }),
+        messages: requestMessages,
        stream: false
      },
      model
@@ -51,13 +53,15 @@ export async function createQuestionGuide({
  const start = answer.indexOf('[');
  const end = answer.lastIndexOf(']');

-  const tokens = await countGptMessagesTokens(concatMessages);
+  const inputTokens = await countGptMessagesTokens(requestMessages);
+  const outputTokens = await countPromptTokens(answer);

  if (start === -1 || end === -1) {
    addLog.warn('Create question guide error', { answer });
    return {
      result: [],
-      tokens: 0
+      inputTokens: 0,
+      outputTokens: 0
    };
  }

@@ -69,14 +73,16 @@ export async function createQuestionGuide({
  try {
    return {
      result: json5.parse(jsonStr),
-      tokens
+      inputTokens,
+      outputTokens
    };
  } catch (error) {
    console.log(error);

    return {
      result: [],
-      tokens: 0
+      inputTokens: 0,
+      outputTokens: 0
    };
  }
 }
--- a/packages/service/core/ai/functions/queryExtension.ts
+++ b/packages/service/core/ai/functions/queryExtension.ts
@@ -1,7 +1,7 @@
 import { replaceVariable } from '@fastgpt/global/common/string/tools';
 import { createChatCompletion } from '../config';
 import { ChatItemType } from '@fastgpt/global/core/chat/type';
-import { countGptMessagesTokens } from '../../../common/string/tiktoken/index';
+import { countGptMessagesTokens, countPromptTokens } from '../../../common/string/tiktoken/index';
 import { chatValue2RuntimePrompt } from '@fastgpt/global/core/chat/adapt';
 import { getLLMModel } from '../model';
 import { llmCompletionsBodyFormat } from '../utils';
@@ -121,7 +121,8 @@ export const queryExtension = async ({
  rawQuery: string;
  extensionQueries: string[];
  model: string;
-  tokens: number;
+  inputTokens: number;
+  outputTokens: number;
 }> => {
  const systemFewShot = chatBg
    ? `Q: 对话背景。
@@ -166,7 +167,8 @@ A: ${chatBg}
      rawQuery: query,
      extensionQueries: [],
      model,
-      tokens: 0
+      inputTokens: 0,
+      outputTokens: 0
    };
  }

@@ -181,7 +183,8 @@ A: ${chatBg}
      rawQuery: query,
      extensionQueries: Array.isArray(queries) ? queries : [],
      model,
-      tokens: await countGptMessagesTokens(messages)
+      inputTokens: await countGptMessagesTokens(messages),
+      outputTokens: await countPromptTokens(answer)
    };
  } catch (error) {
    addLog.error(`Query extension error`, error);
@@ -189,7 +192,8 @@ A: ${chatBg}
      rawQuery: query,
      extensionQueries: [],
      model,
-      tokens: 0
+      inputTokens: 0,
+      outputTokens: 0
    };
  }
 };
--- a/packages/service/core/ai/model.ts
+++ b/packages/service/core/ai/model.ts
@@ -4,6 +4,7 @@ export const getLLMModel = (model?: string) => {
    global.llmModels[0]
  );
 };
+
 export const getDatasetModel = (model?: string) => {
  return (
    global.llmModels
--- a/packages/service/core/workflow/dispatch/agent/classifyQuestion.ts
+++ b/packages/service/core/workflow/dispatch/agent/classifyQuestion.ts
@@ -1,5 +1,8 @@
 import { chats2GPTMessages } from '@fastgpt/global/core/chat/adapt';
-import { countMessagesTokens } from '../../../../common/string/tiktoken/index';
+import {
+  countGptMessagesTokens,
+  countPromptTokens
+} from '../../../../common/string/tiktoken/index';
 import type { ChatItemType } from '@fastgpt/global/core/chat/type.d';
 import { ChatItemValueTypeEnum, ChatRoleEnum } from '@fastgpt/global/core/chat/constants';
 import { createChatCompletion } from '../../../ai/config';
@@ -49,7 +52,7 @@ export const dispatchClassifyQuestion = async (props: Props): Promise<CQResponse

  const chatHistories = getHistories(history, histories);

-  const { arg, tokens } = await completions({
+  const { arg, inputTokens, outputTokens } = await completions({
    ...props,
    histories: chatHistories,
    cqModel
@@ -59,7 +62,8 @@ export const dispatchClassifyQuestion = async (props: Props): Promise<CQResponse

  const { totalPoints, modelName } = formatModelChars2Points({
    model: cqModel.model,
-    tokens,
+    inputTokens: inputTokens,
+    outputTokens: outputTokens,
    modelType: ModelTypeEnum.llm
  });

@@ -72,7 +76,8 @@ export const dispatchClassifyQuestion = async (props: Props): Promise<CQResponse
      totalPoints: externalProvider.openaiAccount?.key ? 0 : totalPoints,
      model: modelName,
      query: userChatInput,
-      tokens,
+      inputTokens: inputTokens,
+      outputTokens: outputTokens,
      cqList: agents,
      cqResult: result.value,
      contextTotalLen: chatHistories.length + 2
@@ -82,7 +87,8 @@ export const dispatchClassifyQuestion = async (props: Props): Promise<CQResponse
        moduleName: name,
        totalPoints: externalProvider.openaiAccount?.key ? 0 : totalPoints,
        model: modelName,
-        tokens
+        inputTokens: inputTokens,
+        outputTokens: outputTokens
      }
    ]
  };
@@ -148,7 +154,8 @@ const completions = async ({
  }

  return {
-    tokens: await countMessagesTokens(messages),
+    inputTokens: await countGptMessagesTokens(requestMessages),
+    outputTokens: await countPromptTokens(answer),
    arg: { type: id }
  };
 };
--- a/packages/service/core/workflow/dispatch/agent/extract.ts
+++ b/packages/service/core/workflow/dispatch/agent/extract.ts
@@ -3,7 +3,8 @@ import { filterGPTMessageByMaxTokens, loadRequestMessages } from '../../../chat/
 import type { ChatItemType } from '@fastgpt/global/core/chat/type.d';
 import {
  countMessagesTokens,
-  countGptMessagesTokens
+  countGptMessagesTokens,
+  countPromptTokens
 } from '../../../../common/string/tiktoken/index';
 import { ChatItemValueTypeEnum, ChatRoleEnum } from '@fastgpt/global/core/chat/constants';
 import { createChatCompletion } from '../../../ai/config';
@@ -59,7 +60,7 @@ export async function dispatchContentExtract(props: Props): Promise<Response> {
  const extractModel = getLLMModel(model);
  const chatHistories = getHistories(history, histories);

-  const { arg, tokens } = await (async () => {
+  const { arg, inputTokens, outputTokens } = await (async () => {
    if (extractModel.toolChoice) {
      return toolChoice({
        ...props,
@@ -114,7 +115,8 @@ export async function dispatchContentExtract(props: Props): Promise<Response> {

  const { totalPoints, modelName } = formatModelChars2Points({
    model: extractModel.model,
-    tokens,
+    inputTokens: inputTokens,
+    outputTokens: outputTokens,
    modelType: ModelTypeEnum.llm
  });

@@ -126,7 +128,8 @@ export async function dispatchContentExtract(props: Props): Promise<Response> {
      totalPoints: externalProvider.openaiAccount?.key ? 0 : totalPoints,
      model: modelName,
      query: content,
-      tokens,
+      inputTokens,
+      outputTokens,
      extractDescription: description,
      extractResult: arg,
      contextTotalLen: chatHistories.length + 2
@@ -136,7 +139,8 @@ export async function dispatchContentExtract(props: Props): Promise<Response> {
        moduleName: name,
        totalPoints: externalProvider.openaiAccount?.key ? 0 : totalPoints,
        model: modelName,
-        tokens
+        inputTokens,
+        outputTokens
      }
    ]
  };
@@ -249,15 +253,18 @@ const toolChoice = async (props: ActionProps) => {
    }
  })();

-  const completeMessages: ChatCompletionMessageParam[] = [
-    ...filterMessages,
+  const AIMessages: ChatCompletionMessageParam[] = [
    {
      role: ChatCompletionRequestMessageRoleEnum.Assistant,
      tool_calls: response.choices?.[0]?.message?.tool_calls
    }
  ];
+
+  const inputTokens = await countGptMessagesTokens(filterMessages, tools);
+  const outputTokens = await countGptMessagesTokens(AIMessages);
  return {
-    tokens: await countGptMessagesTokens(completeMessages, tools),
+    inputTokens,
+    outputTokens,
    arg
  };
 };
@@ -286,17 +293,21 @@ const functionCall = async (props: ActionProps) => {

  try {
    const arg = JSON.parse(response?.choices?.[0]?.message?.function_call?.arguments || '');
-    const completeMessages: ChatCompletionMessageParam[] = [
-      ...filterMessages,
+
+    const AIMessages: ChatCompletionMessageParam[] = [
      {
        role: ChatCompletionRequestMessageRoleEnum.Assistant,
        function_call: response.choices?.[0]?.message?.function_call
      }
    ];

+    const inputTokens = await countGptMessagesTokens(filterMessages, undefined, functions);
+    const outputTokens = await countGptMessagesTokens(AIMessages);
+
    return {
      arg,
-      tokens: await countGptMessagesTokens(completeMessages, undefined, functions)
+      inputTokens,
+      outputTokens
    };
  } catch (error) {
    console.log(response.choices?.[0]?.message);
@@ -305,7 +316,8 @@ const functionCall = async (props: ActionProps) => {

    return {
      arg: {},
-      tokens: 0
+      inputTokens: 0,
+      outputTokens: 0
    };
  }
 };
@@ -370,7 +382,8 @@ Human: ${content}`
  if (!jsonStr) {
    return {
      rawResponse: answer,
-      tokens: await countMessagesTokens(messages),
+      inputTokens: await countMessagesTokens(messages),
+      outputTokens: await countPromptTokens(answer),
      arg: {}
    };
  }
@@ -378,7 +391,8 @@ Human: ${content}`
  try {
    return {
      rawResponse: answer,
-      tokens: await countMessagesTokens(messages),
+      inputTokens: await countMessagesTokens(messages),
+      outputTokens: await countPromptTokens(answer),
      arg: json5.parse(jsonStr) as Record<string, any>
    };
  } catch (error) {
@@ -386,7 +400,8 @@ Human: ${content}`
    console.log(error);
    return {
      rawResponse: answer,
-      tokens: await countMessagesTokens(messages),
+      inputTokens: await countMessagesTokens(messages),
+      outputTokens: await countPromptTokens(answer),
      arg: {}
    };
  }
--- a/packages/service/core/workflow/dispatch/agent/runTool/functionCall.ts
+++ b/packages/service/core/workflow/dispatch/agent/runTool/functionCall.ts
@@ -109,7 +109,8 @@ export const runToolWithFunctionCall = async (

      return {
        dispatchFlowResponse: [toolRunResponse],
-        toolNodeTokens: 0,
+        toolNodeInputTokens: 0,
+        toolNodeOutputTokens: 0,
        completeMessages: requestMessages,
        assistantResponses: toolRunResponse.assistantResponses,
        runTimes: toolRunResponse.runTimes,
@@ -126,7 +127,8 @@ export const runToolWithFunctionCall = async (
      },
      {
        dispatchFlowResponse: [toolRunResponse],
-        toolNodeTokens: 0,
+        toolNodeInputTokens: 0,
+        toolNodeOutputTokens: 0,
        assistantResponses: toolRunResponse.assistantResponses,
        runTimes: toolRunResponse.runTimes
      }
@@ -340,7 +342,9 @@ export const runToolWithFunctionCall = async (
      assistantToolMsgParams
    ] as ChatCompletionMessageParam[];
    // Only toolCall tokens are counted here, Tool response tokens count towards the next reply
-    const tokens = await countGptMessagesTokens(concatToolMessages, undefined, functions);
+    // const tokens = await countGptMessagesTokens(concatToolMessages, undefined, functions);
+    const inputTokens = await countGptMessagesTokens(requestMessages, undefined, functions);
+    const outputTokens = await countGptMessagesTokens([assistantToolMsgParams]);
    /* 
      ...
      user
@@ -375,7 +379,12 @@ export const runToolWithFunctionCall = async (
    const runTimes =
      (response?.runTimes || 0) +
      flatToolsResponseData.reduce((sum, item) => sum + item.runTimes, 0);
-    const toolNodeTokens = response?.toolNodeTokens ? response.toolNodeTokens + tokens : tokens;
+    const toolNodeInputTokens = response?.toolNodeInputTokens
+      ? response.toolNodeInputTokens + inputTokens
+      : inputTokens;
+    const toolNodeOutputTokens = response?.toolNodeOutputTokens
+      ? response.toolNodeOutputTokens + outputTokens
+      : outputTokens;

    // Check stop signal
    const hasStopSignal = flatToolsResponseData.some(
@@ -408,7 +417,8 @@ export const runToolWithFunctionCall = async (

      return {
        dispatchFlowResponse,
-        toolNodeTokens,
+        toolNodeInputTokens,
+        toolNodeOutputTokens,
        completeMessages,
        assistantResponses: toolNodeAssistants,
        runTimes,
@@ -423,7 +433,8 @@ export const runToolWithFunctionCall = async (
      },
      {
        dispatchFlowResponse,
-        toolNodeTokens,
+        toolNodeInputTokens,
+        toolNodeOutputTokens,
        assistantResponses: toolNodeAssistants,
        runTimes
      }
@@ -435,7 +446,8 @@ export const runToolWithFunctionCall = async (
      content: answer
    };
    const completeMessages = filterMessages.concat(gptAssistantResponse);
-    const tokens = await countGptMessagesTokens(completeMessages, undefined, functions);
+    const inputTokens = await countGptMessagesTokens(requestMessages, undefined, functions);
+    const outputTokens = await countGptMessagesTokens([gptAssistantResponse]);
    // console.log(tokens, 'response token');

    // concat tool assistant
@@ -443,7 +455,12 @@ export const runToolWithFunctionCall = async (

    return {
      dispatchFlowResponse: response?.dispatchFlowResponse || [],
-      toolNodeTokens: response?.toolNodeTokens ? response.toolNodeTokens + tokens : tokens,
+      toolNodeInputTokens: response?.toolNodeInputTokens
+        ? response.toolNodeInputTokens + inputTokens
+        : inputTokens,
+      toolNodeOutputTokens: response?.toolNodeOutputTokens
+        ? response.toolNodeOutputTokens + outputTokens
+        : outputTokens,
      completeMessages,
      assistantResponses: [...assistantResponses, ...toolNodeAssistant.value],
      runTimes: (response?.runTimes || 0) + 1
--- a/packages/service/core/workflow/dispatch/agent/runTool/index.ts
+++ b/packages/service/core/workflow/dispatch/agent/runTool/index.ts
@@ -165,6 +165,8 @@ export const dispatchRunTools = async (props: DispatchToolModuleProps): Promise<
    toolWorkflowInteractiveResponse,
    dispatchFlowResponse, // tool flow response
    toolNodeTokens,
+    toolNodeInputTokens,
+    toolNodeOutputTokens,
    completeMessages = [], // The actual message sent to AI(just save text)
    assistantResponses = [], // FastGPT system store assistant.value response
    runTimes
@@ -225,7 +227,8 @@ export const dispatchRunTools = async (props: DispatchToolModuleProps): Promise<

  const { totalPoints, modelName } = formatModelChars2Points({
    model,
-    tokens: toolNodeTokens,
+    inputTokens: toolNodeInputTokens,
+    outputTokens: toolNodeOutputTokens,
    modelType: ModelTypeEnum.llm
  });
  const toolAIUsage = externalProvider.openaiAccount?.key ? 0 : totalPoints;
@@ -255,6 +258,8 @@ export const dispatchRunTools = async (props: DispatchToolModuleProps): Promise<
      // 展示的积分消耗
      totalPoints: totalPointsUsage,
      toolCallTokens: toolNodeTokens,
+      toolCallInputTokens: toolNodeInputTokens,
+      toolCallOutputTokens: toolNodeOutputTokens,
      childTotalPoints: flatUsages.reduce((sum, item) => sum + item.totalPoints, 0),
      model: modelName,
      query: userChatInput,
@@ -270,9 +275,10 @@ export const dispatchRunTools = async (props: DispatchToolModuleProps): Promise<
      // 工具调用本身的积分消耗
      {
        moduleName: name,
-        totalPoints: toolAIUsage,
        model: modelName,
-        tokens: toolNodeTokens
+        totalPoints: toolAIUsage,
+        inputTokens: toolNodeInputTokens,
+        outputTokens: toolNodeOutputTokens
      },
      // 工具的消耗
      ...flatUsages
--- a/packages/service/core/workflow/dispatch/agent/runTool/promptCall.ts
+++ b/packages/service/core/workflow/dispatch/agent/runTool/promptCall.ts
@@ -115,7 +115,8 @@ export const runToolWithPromptCall = async (

      return {
        dispatchFlowResponse: [toolRunResponse],
-        toolNodeTokens: 0,
+        toolNodeInputTokens: 0,
+        toolNodeOutputTokens: 0,
        completeMessages: concatMessages,
        assistantResponses: toolRunResponse.assistantResponses,
        runTimes: toolRunResponse.runTimes,
@@ -131,7 +132,8 @@ export const runToolWithPromptCall = async (
      },
      {
        dispatchFlowResponse: [toolRunResponse],
-        toolNodeTokens: 0,
+        toolNodeInputTokens: 0,
+        toolNodeOutputTokens: 0,
        assistantResponses: toolRunResponse.assistantResponses,
        runTimes: toolRunResponse.runTimes
      }
@@ -286,15 +288,20 @@ export const runToolWithPromptCall = async (
      content: replaceAnswer
    };
    const completeMessages = filterMessages.concat(gptAssistantResponse);
-    const tokens = await countGptMessagesTokens(completeMessages, undefined);
-    // console.log(tokens, 'response token');
+    const inputTokens = await countGptMessagesTokens(requestMessages);
+    const outputTokens = await countGptMessagesTokens([gptAssistantResponse]);

    // concat tool assistant
    const toolNodeAssistant = GPTMessages2Chats([gptAssistantResponse])[0] as AIChatItemType;

    return {
      dispatchFlowResponse: response?.dispatchFlowResponse || [],
-      toolNodeTokens: response?.toolNodeTokens ? response.toolNodeTokens + tokens : tokens,
+      toolNodeInputTokens: response?.toolNodeInputTokens
+        ? response.toolNodeInputTokens + inputTokens
+        : inputTokens,
+      toolNodeOutputTokens: response?.toolNodeOutputTokens
+        ? response.toolNodeOutputTokens + outputTokens
+        : outputTokens,
      completeMessages,
      assistantResponses: [...assistantResponses, ...toolNodeAssistant.value],
      runTimes: (response?.runTimes || 0) + 1
@@ -366,17 +373,9 @@ export const runToolWithPromptCall = async (
    function_call: toolJson
  };

-  /* 
-    ...
-    user
-    assistant: tool data
-  */
-  const concatToolMessages = [
-    ...requestMessages,
-    assistantToolMsgParams
-  ] as ChatCompletionMessageParam[];
  // Only toolCall tokens are counted here, Tool response tokens count towards the next reply
-  const tokens = await countGptMessagesTokens(concatToolMessages, undefined);
+  const inputTokens = await countGptMessagesTokens(requestMessages);
+  const outputTokens = await countGptMessagesTokens([assistantToolMsgParams]);

  /* 
    ...
@@ -437,7 +436,12 @@ ANSWER: `;
  }

  const runTimes = (response?.runTimes || 0) + toolsRunResponse.toolResponse.runTimes;
-  const toolNodeTokens = response?.toolNodeTokens ? response.toolNodeTokens + tokens : tokens;
+  const toolNodeInputTokens = response?.toolNodeInputTokens
+    ? response.toolNodeInputTokens + inputTokens
+    : inputTokens;
+  const toolNodeOutputTokens = response?.toolNodeOutputTokens
+    ? response.toolNodeOutputTokens + outputTokens
+    : outputTokens;

  // Check stop signal
  const hasStopSignal = toolsRunResponse.toolResponse.flowResponses.some((item) => !!item.toolStop);
@@ -460,7 +464,8 @@ ANSWER: `;

    return {
      dispatchFlowResponse,
-      toolNodeTokens,
+      toolNodeInputTokens,
+      toolNodeOutputTokens,
      completeMessages: filterMessages,
      assistantResponses: toolNodeAssistants,
      runTimes,
@@ -475,7 +480,8 @@ ANSWER: `;
    },
    {
      dispatchFlowResponse,
-      toolNodeTokens,
+      toolNodeInputTokens,
+      toolNodeOutputTokens,
      assistantResponses: toolNodeAssistants,
      runTimes
    }
--- a/packages/service/core/workflow/dispatch/agent/runTool/toolChoice.ts
+++ b/packages/service/core/workflow/dispatch/agent/runTool/toolChoice.ts
@@ -158,7 +158,8 @@ export const runToolWithToolChoice = async (

      return {
        dispatchFlowResponse: [toolRunResponse],
-        toolNodeTokens: 0,
+        toolNodeInputTokens: 0,
+        toolNodeOutputTokens: 0,
        completeMessages: requestMessages,
        assistantResponses: toolRunResponse.assistantResponses,
        runTimes: toolRunResponse.runTimes,
@@ -176,7 +177,8 @@ export const runToolWithToolChoice = async (
      },
      {
        dispatchFlowResponse: [toolRunResponse],
-        toolNodeTokens: 0,
+        toolNodeInputTokens: 0,
+        toolNodeOutputTokens: 0,
        assistantResponses: toolRunResponse.assistantResponses,
        runTimes: toolRunResponse.runTimes
      }
@@ -428,7 +430,9 @@ export const runToolWithToolChoice = async (
    ] as ChatCompletionMessageParam[];

    // Only toolCall tokens are counted here, Tool response tokens count towards the next reply
-    const tokens = await countGptMessagesTokens(concatToolMessages, tools);
+    const inputTokens = await countGptMessagesTokens(requestMessages, tools);
+    const outputTokens = await countGptMessagesTokens(assistantToolMsgParams);
+
    /* 
        ...
        user
@@ -463,7 +467,10 @@ export const runToolWithToolChoice = async (
    const runTimes =
      (response?.runTimes || 0) +
      flatToolsResponseData.reduce((sum, item) => sum + item.runTimes, 0);
-    const toolNodeTokens = response ? response.toolNodeTokens + tokens : tokens;
+    const toolNodeInputTokens = response ? response.toolNodeInputTokens + inputTokens : inputTokens;
+    const toolNodeOutputTokens = response
+      ? response.toolNodeOutputTokens + outputTokens
+      : outputTokens;

    // Check stop signal
    const hasStopSignal = flatToolsResponseData.some(
@@ -496,7 +503,8 @@ export const runToolWithToolChoice = async (

      return {
        dispatchFlowResponse,
-        toolNodeTokens,
+        toolNodeInputTokens,
+        toolNodeOutputTokens,
        completeMessages,
        assistantResponses: toolNodeAssistants,
        runTimes,
@@ -512,7 +520,8 @@ export const runToolWithToolChoice = async (
      },
      {
        dispatchFlowResponse,
-        toolNodeTokens,
+        toolNodeInputTokens,
+        toolNodeOutputTokens,
        assistantResponses: toolNodeAssistants,
        runTimes
      }
@@ -524,14 +533,17 @@ export const runToolWithToolChoice = async (
      content: answer
    };
    const completeMessages = filterMessages.concat(gptAssistantResponse);
-    const tokens = await countGptMessagesTokens(completeMessages, tools);
+    const inputTokens = await countGptMessagesTokens(requestMessages, tools);
+    const outputTokens = await countGptMessagesTokens([gptAssistantResponse]);

    // concat tool assistant
    const toolNodeAssistant = GPTMessages2Chats([gptAssistantResponse])[0] as AIChatItemType;

    return {
      dispatchFlowResponse: response?.dispatchFlowResponse || [],
-      toolNodeTokens: response ? response.toolNodeTokens + tokens : tokens,
+      toolNodeInputTokens: response ? response.toolNodeInputTokens + inputTokens : inputTokens,
+      toolNodeOutputTokens: response ? response.toolNodeOutputTokens + outputTokens : outputTokens,
+
      completeMessages,
      assistantResponses: [...assistantResponses, ...toolNodeAssistant.value],
      runTimes: (response?.runTimes || 0) + 1
@@ -578,7 +590,8 @@ async function streamResponse({
          text: content
        })
      });
-    } else if (responseChoice?.tool_calls?.[0]) {
+    }
+    if (responseChoice?.tool_calls?.[0]) {
      const toolCall: ChatCompletionMessageToolCall = responseChoice.tool_calls[0];
      // In a stream response, only one tool is returned at a time.  If have id, description is executing a tool
      if (toolCall.id || callingTool) {
--- a/packages/service/core/workflow/dispatch/agent/runTool/type.d.ts
+++ b/packages/service/core/workflow/dispatch/agent/runTool/type.d.ts
@@ -31,7 +31,9 @@ export type DispatchToolModuleProps = ModuleDispatchProps<{

 export type RunToolResponse = {
  dispatchFlowResponse: DispatchFlowResponse[];
-  toolNodeTokens: number;
+  toolNodeTokens?: number; // deprecated
+  toolNodeInputTokens: number;
+  toolNodeOutputTokens: number;
  completeMessages?: ChatCompletionMessageParam[];
  assistantResponses?: AIChatItemValueItemType[];
  toolWorkflowInteractiveResponse?: WorkflowInteractiveResponseType;
--- a/packages/service/core/workflow/dispatch/chat/oneapi.ts
+++ b/packages/service/core/workflow/dispatch/chat/oneapi.ts
@@ -5,13 +5,17 @@ import { ChatRoleEnum } from '@fastgpt/global/core/chat/constants';
 import { SseResponseEventEnum } from '@fastgpt/global/core/workflow/runtime/constants';
 import { textAdaptGptResponse } from '@fastgpt/global/core/workflow/runtime/utils';
 import { createChatCompletion } from '../../../ai/config';
-import type { ChatCompletion, StreamChatType } from '@fastgpt/global/core/ai/type.d';
+import type {
+  ChatCompletion,
+  ChatCompletionMessageParam,
+  StreamChatType
+} from '@fastgpt/global/core/ai/type.d';
 import { formatModelChars2Points } from '../../../../support/wallet/usage/utils';
 import type { LLMModelItemType } from '@fastgpt/global/core/ai/model.d';
 import { postTextCensor } from '../../../../common/api/requestPlusApi';
 import { ChatCompletionRequestMessageRoleEnum } from '@fastgpt/global/core/ai/constants';
 import type { DispatchNodeResultType } from '@fastgpt/global/core/workflow/runtime/type';
-import { countMessagesTokens } from '../../../../common/string/tiktoken/index';
+import { countGptMessagesTokens } from '../../../../common/string/tiktoken/index';
 import {
  chats2GPTMessages,
  chatValue2RuntimePrompt,
@@ -214,16 +218,23 @@ export const dispatchChatCompletion = async (props: ChatProps): Promise<ChatResp
    return Promise.reject(getEmptyResponseTip());
  }

-  const completeMessages = requestMessages.concat({
-    role: ChatCompletionRequestMessageRoleEnum.Assistant,
-    content: answerText
-  });
+  const AIMessages: ChatCompletionMessageParam[] = [
+    {
+      role: ChatCompletionRequestMessageRoleEnum.Assistant,
+      content: answerText
+    }
+  ];
+
+  const completeMessages = [...requestMessages, ...AIMessages];
  const chatCompleteMessages = GPTMessages2Chats(completeMessages);

-  const tokens = await countMessagesTokens(chatCompleteMessages);
+  const inputTokens = await countGptMessagesTokens(requestMessages);
+  const outputTokens = await countGptMessagesTokens(AIMessages);
+
  const { totalPoints, modelName } = formatModelChars2Points({
    model,
-    tokens,
+    inputTokens,
+    outputTokens,
    modelType: ModelTypeEnum.llm
  });

@@ -232,7 +243,9 @@ export const dispatchChatCompletion = async (props: ChatProps): Promise<ChatResp
    [DispatchNodeResponseKeyEnum.nodeResponse]: {
      totalPoints: externalProvider.openaiAccount?.key ? 0 : totalPoints,
      model: modelName,
-      tokens,
+      tokens: inputTokens + outputTokens,
+      inputTokens: inputTokens,
+      outputTokens: outputTokens,
      query: `${userChatInput}`,
      maxToken: max_tokens,
      historyPreview: getHistoryPreview(
@@ -247,7 +260,8 @@ export const dispatchChatCompletion = async (props: ChatProps): Promise<ChatResp
        moduleName: name,
        totalPoints: externalProvider.openaiAccount?.key ? 0 : totalPoints,
        model: modelName,
-        tokens
+        inputTokens: inputTokens,
+        outputTokens: outputTokens
      }
    ],
    [DispatchNodeResponseKeyEnum.toolResponses]: answerText,
--- a/packages/service/core/workflow/dispatch/dataset/search.ts
+++ b/packages/service/core/workflow/dispatch/dataset/search.ts
@@ -120,14 +120,14 @@ export async function dispatchDatasetSearch(
  // vector
  const { totalPoints, modelName } = formatModelChars2Points({
    model: vectorModel.model,
-    tokens,
+    inputTokens: tokens,
    modelType: ModelTypeEnum.vector
  });
  const responseData: DispatchNodeResponseType & { totalPoints: number } = {
    totalPoints,
    query: concatQueries.join('\n'),
    model: modelName,
-    tokens,
+    inputTokens: tokens,
    similarity: usingSimilarityFilter ? similarity : undefined,
    limit,
    searchMode,
@@ -139,19 +139,21 @@ export async function dispatchDatasetSearch(
      totalPoints,
      moduleName: node.name,
      model: modelName,
-      tokens
+      inputTokens: tokens
    }
  ];

  if (aiExtensionResult) {
    const { totalPoints, modelName } = formatModelChars2Points({
      model: aiExtensionResult.model,
-      tokens: aiExtensionResult.tokens,
+      inputTokens: aiExtensionResult.inputTokens,
+      outputTokens: aiExtensionResult.outputTokens,
      modelType: ModelTypeEnum.llm
    });

    responseData.totalPoints += totalPoints;
-    responseData.tokens = aiExtensionResult.tokens;
+    responseData.inputTokens = aiExtensionResult.inputTokens;
+    responseData.outputTokens = aiExtensionResult.outputTokens;
    responseData.extensionModel = modelName;
    responseData.extensionResult =
      aiExtensionResult.extensionQueries?.join('\n') ||
@@ -161,7 +163,8 @@ export async function dispatchDatasetSearch(
      totalPoints,
      moduleName: 'core.module.template.Query extension',
      model: modelName,
-      tokens: aiExtensionResult.tokens
+      inputTokens: aiExtensionResult.inputTokens,
+      outputTokens: aiExtensionResult.outputTokens
    });
  }

--- a/packages/service/core/workflow/dispatch/plugin/run.ts
+++ b/packages/service/core/workflow/dispatch/plugin/run.ts
@@ -130,8 +130,7 @@ export const dispatchRunPlugin = async (props: RunPluginProps): Promise<RunPlugi
    [DispatchNodeResponseKeyEnum.nodeDispatchUsages]: [
      {
        moduleName: plugin.name,
-        totalPoints: usagePoints,
-        tokens: 0
+        totalPoints: usagePoints
      }
    ],
    [DispatchNodeResponseKeyEnum.toolResponses]: output?.pluginOutput
--- a/packages/service/core/workflow/dispatch/plugin/runApp.ts
+++ b/packages/service/core/workflow/dispatch/plugin/runApp.ts
@@ -153,8 +153,7 @@ export const dispatchRunAppNode = async (props: Props): Promise<Response> => {
    [DispatchNodeResponseKeyEnum.nodeDispatchUsages]: [
      {
        moduleName: appData.name,
-        totalPoints: usagePoints,
-        tokens: 0
+        totalPoints: usagePoints
      }
    ],
    [DispatchNodeResponseKeyEnum.toolResponses]: text,
--- a/packages/service/core/workflow/dispatch/tools/queryExternsion.ts
+++ b/packages/service/core/workflow/dispatch/tools/queryExternsion.ts
@@ -31,7 +31,7 @@ export const dispatchQueryExtension = async ({
  const queryExtensionModel = getLLMModel(model);
  const chatHistories = getHistories(history, histories);

-  const { extensionQueries, tokens } = await queryExtension({
+  const { extensionQueries, inputTokens, outputTokens } = await queryExtension({
    chatBg: systemPrompt,
    query: userChatInput,
    histories: chatHistories,
@@ -42,7 +42,8 @@ export const dispatchQueryExtension = async ({

  const { totalPoints, modelName } = formatModelChars2Points({
    model: queryExtensionModel.model,
-    tokens,
+    inputTokens,
+    outputTokens,
    modelType: ModelTypeEnum.llm
  });

@@ -59,7 +60,8 @@ export const dispatchQueryExtension = async ({
    [DispatchNodeResponseKeyEnum.nodeResponse]: {
      totalPoints,
      model: modelName,
-      tokens,
+      inputTokens,
+      outputTokens,
      query: userChatInput,
      textOutput: JSON.stringify(filterSameQueries)
    },
@@ -68,7 +70,8 @@ export const dispatchQueryExtension = async ({
        moduleName: node.name,
        totalPoints,
        model: modelName,
-        tokens
+        inputTokens,
+        outputTokens
      }
    ],
    [NodeOutputKeyEnum.text]: JSON.stringify(filterSameQueries)