kefu/knowledge/controller/gpt.go

package controller

import (
	"encoding/json"
	"errors"
	"fmt"
	"github.com/gin-gonic/gin"
	"github.com/tidwall/gjson"
	"io"
	"knowledge/utils"
	"log"
	"net/http"
	"net/url"
	"os"
	"strconv"
)

func PostChatGPT(c *gin.Context) {
	// 将响应头中的Content-Type设置为text/plain，表示响应内容为文本
	c.Header("Content-Type", "text/html;charset=utf-8;")
	// 关闭输出缓冲，使得每次写入的数据能够立即发送给客户端
	f, ok := c.Writer.(http.Flusher)

	keywords, _ := url.QueryUnescape(c.Query("keywords"))
	if keywords == "" {
		keywords = c.PostForm("keywords")
	}
	system, _ := url.QueryUnescape(c.Query("system"))
	if system == "" {
		system = c.PostForm("system")
	}
	if system == "" {
		system = "你正在扮演知识库AI机器人"
	}
	if keywords == "" {
		c.Writer.Write([]byte("请求参数为空"))
		return
	}
	//历史记录
	history := c.PostForm("history")
	//匹配不到
	prompt := c.PostForm("prompt")

	if prompt != "" {
		prompt = fmt.Sprintf("对于与知识信息无关的问题，你应拒绝并告知用户“%s”", prompt)
	}
	openaiKey := os.Getenv("OPENAI_KEY")
	openaiUrl := os.Getenv("OPENAI_API_BASE")
	//gpt url
	gptUrl := c.PostForm("gptUrl")
	if gptUrl != "" {
		openaiUrl = gptUrl
	}
	//gpt secret
	gptSecret := c.PostForm("gptSecret")
	if gptSecret != "" {
		openaiKey = gptSecret
	}
	collectName := c.Param("collectName")
	log.Println(openaiUrl, openaiKey)
	gpt := utils.NewChatGptTool(openaiUrl, openaiKey)
	message, err, urls := MakePrompt(openaiUrl, openaiKey, collectName, keywords, system, history, prompt)
	if err != nil {
		c.Writer.Write([]byte(err.Error()))
		return
	}
	log.Printf("请求openai：%+v\n", message)
	stream, err := gpt.ChatGPT3Dot5TurboStream(message)
	if err != nil {
		log.Println("gpt3 stream error: ", err)
		return
	}
	if !ok {
		c.AbortWithStatus(http.StatusInternalServerError)
		return
	}
	for {
		response, err := stream.Recv()
		if errors.Is(err, io.EOF) {
			log.Println("\nStream finished", err)
			break
		} else if err != nil {
			log.Printf("\nStream error: %v,%v\n", err, response)
			break
		} else {
			data := response.Choices[0].Delta.Content
			log.Printf(data)
			c.Writer.Write([]byte(data))
			f.Flush()
		}
	}
	isShowRefer := os.Getenv("IS_SHOW_REFER")
	if len(urls) != 0 && isShowRefer == "yes" {
		log.Println(urls)
		urls = utils.RemoveRepByMap(urls)
		c.Writer.Write([]byte("\r\n\r\n参考信息"))
		f.Flush()
		i := 1
		for _, url := range urls {
			utlHtml := fmt.Sprintf("\n\n%d. %s", i, url)
			c.Writer.Write([]byte(utlHtml))
			f.Flush()
			i++
		}
	}

	stream.Close()
}

//组织Prompt
func MakePrompt(openaiUrl, openaiKey, collectionName, keywords, system, history, promptMessage string) ([]utils.Gpt3Dot5Message, error, []string) {
	gpt := utils.NewChatGptTool(openaiUrl, openaiKey)
	urls := make([]string, 0)

	response, err := gpt.GetEmbedding(keywords, "text-embedding-ada-002")
	if err != nil {
		return nil, err, urls
	}
	var embeddingResponse utils.EmbeddingResponse
	json.Unmarshal([]byte(response), &embeddingResponse)

	params := map[string]interface{}{"exact": false, "hnsw_ef": 128}
	if len(embeddingResponse.Data) == 0 {
		return nil, errors.New(response), urls
	}
	vector := embeddingResponse.Data[0].Embedding
	limit := 3
	points, _ := utils.SearchPoints(collectionName, params, vector, limit)
	log.Printf("相似搜索结果：%s\n", string(points))
	result := gjson.Get(string(points), "result").Array()
	message := make([]utils.Gpt3Dot5Message, 0)
	content := ""
	line := ""
	referScoreStr := os.Getenv("SHOW_REFER_SCORE")
	referScore := 0.85
	if referScoreStr != "" {
		referScore, _ = strconv.ParseFloat(referScoreStr, 64)
	}
	for key, row := range result {
		key++
		line = fmt.Sprintf("%d. %s\n", key, row.Get("payload.text").String())
		url := row.Get("payload.url").String()
		score := row.Get("score").Float()
		if score >= referScore && url != "" && key <= 3 {
			urls = append(urls, url)
		}
		arr := []rune(line)
		if len(arr) > 500 {
			line = string(arr[:500])
		}
		//message = append(message, utils.Gpt3Dot5Message{
		//	Role:    "assistant",
		//	Content: line,
		//})
		content += line
	}
	if content == "" {
		content = "我是知识库AI机器人\n"
	}
	message = append(message, utils.Gpt3Dot5Message{
		Role: "system",
		//Content: fmt.Sprintf("%s，我会向你提问题，我提供的上下文信息是：\n\n%s\n", system, content),
		Content: fmt.Sprintf("%s，%s\n我提供的知识信息是：\n\n%s\n", system, promptMessage, content),
		//Content: fmt.Sprintf("%s，我会向你提问题，你只能基于知识信息回答问题，我提供的知识信息是：\n\n%s\n\n你必须根据自己的知识信息总结归纳后回答问题，不要写解释，不要写具体代码，对于与知识信息无关的问题或者不理解的问题等，你应拒绝并告知用户“未查询到相关信息，请提供详细的问题信息。”", system, content),
		//Content: fmt.Sprintf("你现在扮演知识库AI机器人。请严格根据提供的参考信息总结归纳后回答问题。对于与参考信息无关的问题或者不理解的问题等，你应拒绝并告知用户“未查询到相关信息，请提供详细的问题信息。”避免引用任何当前或过去的政治人物或事件，以及可能引起争议或分裂的历史人物或事件。",
	})
	var historySlice = make([]utils.Gpt3Dot5Message, 0)
	err = json.Unmarshal([]byte(history), &historySlice)
	if err == nil {
		message = append(message, historySlice...)
	} else {
		//message = append(message, utils.Gpt3Dot5Message{
		//	Role: "user",
		//	//Content: keywords,
		//	Content: fmt.Sprintf("我提供的知识信息是：\n\n%s\n\n你必须根据提供的知识信息总结归纳后回答问题，不要写解释，不要写具体代码，对于与知识信息无关的问题或者不理解的问题等，你应拒绝并告知用户“未查询到相关信息，请提供详细的问题信息。”，我的问题是：%s", content, keywords),
		//})
	}
	//userQuestion := keywords
	//if promptMessage != "" {
	userQuestion := fmt.Sprintf("%s，%s\n我的问题是：'''\n%s\n'''", system, promptMessage, keywords)
	//}

	message = append(message, utils.Gpt3Dot5Message{
		Role:    "user",
		Content: userQuestion,
	})
	return message, nil, urls
}