FastGPT/packages/service/core/ai/utils.ts

import { LLMModelItemType } from '@fastgpt/global/core/ai/model.d';
import {
  ChatCompletionCreateParamsNonStreaming,
  ChatCompletionCreateParamsStreaming,
  CompletionFinishReason,
  StreamChatType
} from '@fastgpt/global/core/ai/type';
import { getLLMModel } from './model';

/* 
  Count response max token
*/
export const computedMaxToken = ({
  maxToken,
  model
}: {
  maxToken?: number;
  model: LLMModelItemType;
}) => {
  if (maxToken === undefined) return;

  maxToken = Math.min(maxToken, model.maxResponse);
  return maxToken;
};

// FastGPT temperature range: [0,10], ai temperature:[0,2],{0,1]……
export const computedTemperature = ({
  model,
  temperature
}: {
  model: LLMModelItemType;
  temperature: number;
}) => {
  if (typeof model.maxTemperature !== 'number') return undefined;
  temperature = +(model.maxTemperature * (temperature / 10)).toFixed(2);
  temperature = Math.max(temperature, 0.01);

  return temperature;
};

type CompletionsBodyType =
  | ChatCompletionCreateParamsNonStreaming
  | ChatCompletionCreateParamsStreaming;
type InferCompletionsBody<T> = T extends { stream: true }
  ? ChatCompletionCreateParamsStreaming
  : T extends { stream: false }
    ? ChatCompletionCreateParamsNonStreaming
    : ChatCompletionCreateParamsNonStreaming | ChatCompletionCreateParamsStreaming;

export const llmCompletionsBodyFormat = <T extends CompletionsBodyType>(
  body: T & {
    response_format?: any;
    json_schema?: string;
    stop?: string;
  },
  model: string | LLMModelItemType
): InferCompletionsBody<T> => {
  const modelData = typeof model === 'string' ? getLLMModel(model) : model;
  if (!modelData) {
    return body as unknown as InferCompletionsBody<T>;
  }

  const response_format = body.response_format;
  const json_schema = body.json_schema ?? undefined;
  const stop = body.stop ?? undefined;

  const requestBody: T = {
    ...body,
    model: modelData.model,
    temperature:
      typeof body.temperature === 'number'
        ? computedTemperature({
            model: modelData,
            temperature: body.temperature
          })
        : undefined,
    ...modelData?.defaultConfig,
    response_format: response_format
      ? {
          type: response_format,
          json_schema
        }
      : undefined,
    stop: stop?.split('|')
  };

  // field map
  if (modelData.fieldMap) {
    Object.entries(modelData.fieldMap).forEach(([sourceKey, targetKey]) => {
      // @ts-ignore
      requestBody[targetKey] = body[sourceKey];
      // @ts-ignore
      delete requestBody[sourceKey];
    });
  }

  return requestBody as unknown as InferCompletionsBody<T>;
};

export const llmStreamResponseToAnswerText = async (response: StreamChatType) => {
  let answer = '';
  for await (const part of response) {
    const content = part.choices?.[0]?.delta?.content || '';
    answer += content;
  }
  return parseReasoningContent(answer)[1];
};

// Parse <think></think> tags to think and answer - unstream response
export const parseReasoningContent = (text: string): [string, string] => {
  const regex = /<think>([\s\S]*?)<\/think>/;
  const match = text.match(regex);

  if (!match) {
    return ['', text];
  }

  const thinkContent = match[1].trim();

  // Add answer (remaining text after think tag)
  const answerContent = text.slice(match.index! + match[0].length);

  return [thinkContent, answerContent];
};

// Parse <think></think> tags to think and answer - stream response
export const parseReasoningStreamContent = () => {
  let isInThinkTag: boolean | undefined;

  const startTag = '<think>';
  let startTagBuffer = '';

  const endTag = '</think>';
  let endTagBuffer = '';

  /* 
    parseThinkTag - 只控制是否主动解析 <think></think>，如果接口已经解析了，则不再解析。
  */
  const parsePart = (
    part: {
      choices: {
        delta: {
          content?: string;
          reasoning_content?: string;
        };
        finish_reason?: CompletionFinishReason;
      }[];
    },
    parseThinkTag = false
  ): {
    reasoningContent: string;
    content: string;
    finishReason: CompletionFinishReason;
  } => {
    const content = part.choices?.[0]?.delta?.content || '';
    const finishReason = part.choices?.[0]?.finish_reason || null;

    // @ts-ignore
    const reasoningContent = part.choices?.[0]?.delta?.reasoning_content || '';
    if (reasoningContent || !parseThinkTag) {
      isInThinkTag = false;
      return { reasoningContent, content, finishReason };
    }

    if (!content) {
      return {
        reasoningContent: '',
        content: '',
        finishReason
      };
    }

    // 如果不在 think 标签中，或者有 reasoningContent(接口已解析），则返回 reasoningContent 和 content
    if (isInThinkTag === false) {
      return {
        reasoningContent: '',
        content,
        finishReason
      };
    }

    // 检测是否为 think 标签开头的数据
    if (isInThinkTag === undefined) {
      // Parse content think and answer
      startTagBuffer += content;
      // 太少内容时候，暂时不解析
      if (startTagBuffer.length < startTag.length) {
        return {
          reasoningContent: '',
          content: '',
          finishReason
        };
      }

      if (startTagBuffer.startsWith(startTag)) {
        isInThinkTag = true;
        return {
          reasoningContent: startTagBuffer.slice(startTag.length),
          content: '',
          finishReason
        };
      }

      // 如果未命中 think 标签，则认为不在 think 标签中，返回 buffer 内容作为 content
      isInThinkTag = false;
      return {
        reasoningContent: '',
        content: startTagBuffer,
        finishReason
      };
    }

    // 确认是 think 标签内容，开始返回 think 内容，并实时检测 </think>
    /* 
      检测 </think> 方案。
      存储所有疑似 </think> 的内容，直到检测到完整的 </think> 标签或超出 </think> 长度。
      content 返回值包含以下几种情况:
        abc - 完全未命中尾标签
        abc<th - 命中一部分尾标签
        abc</think> - 完全命中尾标签
        abc</think>abc - 完全命中尾标签
        </think>abc - 完全命中尾标签
        k>abc - 命中一部分尾标签
    */
    // endTagBuffer 专门用来记录疑似尾标签的内容
    if (endTagBuffer) {
      endTagBuffer += content;
      if (endTagBuffer.includes(endTag)) {
        isInThinkTag = false;
        const answer = endTagBuffer.slice(endTag.length);
        return {
          reasoningContent: '',
          content: answer,
          finishReason
        };
      } else if (endTagBuffer.length >= endTag.length) {
        // 缓存内容超出尾标签长度，且仍未命中 </think>，则认为本次猜测 </think> 失败，仍处于 think 阶段。
        const tmp = endTagBuffer;
        endTagBuffer = '';
        return {
          reasoningContent: tmp,
          content: '',
          finishReason
        };
      }
      return {
        reasoningContent: '',
        content: '',
        finishReason
      };
    } else if (content.includes(endTag)) {
      // 返回内容，完整命中</think>，直接结束
      isInThinkTag = false;
      const [think, answer] = content.split(endTag);
      return {
        reasoningContent: think,
        content: answer,
        finishReason
      };
    } else {
      // 无 buffer，且未命中 </think>，开始疑似 </think> 检测。
      for (let i = 1; i < endTag.length; i++) {
        const partialEndTag = endTag.slice(0, i);
        // 命中一部分尾标签
        if (content.endsWith(partialEndTag)) {
          const think = content.slice(0, -partialEndTag.length);
          endTagBuffer += partialEndTag;
          return {
            reasoningContent: think,
            content: '',
            finishReason
          };
        }
      }
    }

    // 完全未命中尾标签，还是 think 阶段。
    return {
      reasoningContent: content,
      content: '',
      finishReason
    };
  };

  const getStartTagBuffer = () => startTagBuffer;

  return {
    parsePart,
    getStartTagBuffer
  };
};