tgbot_misaka_5882f7/assistant.go

package main

import (
	"crypto/md5"
	"encoding/base64"
	"strconv"
	"strings"
	"time"
	"unicode"

	"git.gensokyo.cafe/kkyy/tgbot_misaka_5882f7/openai"
	"git.gensokyo.cafe/kkyy/tgbot_misaka_5882f7/openai/prompts"
	"github.com/samber/lo"
	tele "gopkg.in/telebot.v3"
)

var assistantWritingSign = "\n... 📝"

func matchAssistantConversation(botUsr *tele.User, msg *tele.Message) []*tele.Message {
	// A thread must meet the following conditions to be considered a conversation with the assistant (for now):
	// It has only two participants: the assistant and the user.
	// Or, it has only one participant, which is the user. In this case, it has to be in a private chat, or the assistant must be mentioned if it's in a group chat.
	// No bot commands, images, stickers, or other message types involved.
	participants := map[string]struct{}{}
	hasAssistant := false
	hasUser := false
	assistantMentioned := false
	isPrivateChat := true

	thread, err := getCachedThread(msg)
	if err != nil {
		return nil
	}

	for _, msg := range thread {
		for _, ent := range msg.Entities {
			if ent.Type == tele.EntityCommand {
				return nil
			}
		}

		from := msg.Sender
		chat := msg.Chat
		if from == nil || chat == nil {
			return nil
		}
		participant := strconv.FormatInt(chat.ID, 10) + ":" + strconv.FormatInt(from.ID, 10)
		participants[participant] = struct{}{}

		if chat.Type != tele.ChatPrivate {
			isPrivateChat = false
		}
		if !lo.Contains([]tele.ChatType{tele.ChatPrivate, tele.ChatGroup, tele.ChatSuperGroup}, chat.Type) {
			return nil
		}

		if msg.Text == "" {
			return nil
		}

		if from.ID == botUsr.ID {
			hasAssistant = true
			if strings.HasSuffix(msg.Text, assistantWritingSign) {
				return nil
			}
		} else {
			if from.IsBot {
				return nil
			}
			hasUser = true
			// Only @bot_user_name in the beginning will be counted as "mention"
			mentionStr := "@" + botUsr.Username
			if strings.HasPrefix(strings.TrimSpace(msg.Text), mentionStr) {
				assistantMentioned = true
			}
		}
	}

	if len(participants) > 2 {
		return nil
	}
	if hasAssistant && hasUser {
		return thread
	}
	if hasUser && (isPrivateChat || assistantMentioned) {
		return thread
	}

	return nil
}

type assistantStreamedResponseCb func(text string, finished bool) error

// TODO interrupt response with context.

func assistantStreamedResponse(request openai.ChatRequest, cb assistantStreamedResponseCb) error {
	logger.Debugw("Openai chat request", "req", request)
	ai := openai.NewClient(config.OpenAIApiKey)

	var (
		resp *openai.ChatResponseStream
		err  error
	)
	tries := 2
	for {
		if tries--; tries < 0 {
			return err
		}
		if resp, err = ai.ChatCompletionStream(request); err == nil {
			break
		}
		logger.Warnw("assistant: failed to get response", "error", err)
	}

	var nErrs, nUpdates int
	go func() {
		respBuilder := strings.Builder{}
		minWait := time.After(time.Second)
		for {
			nUpdates += 1

			var (
				nNewChunk        int
				finished         bool
				minWaitSatisfied bool
			)
		Drain:
			for {
				select {
				case chunk, ok := <-resp.Stream:
					if !ok {
						finished = true
						break Drain
					}
					nNewChunk += 1
					respBuilder.WriteString(chunk)
				default:
					if minWaitSatisfied {
						break Drain
					}
					select {
					case <-minWait:
					case <-resp.Done:
					}
					minWaitSatisfied = true
				}
			}

			if nNewChunk == 0 {
				if chunk, ok := <-resp.Stream; !ok {
					finished = true
				} else {
					respBuilder.WriteString(chunk)
				}
			}
			if finished {
				break
			}

			// Renew the timer.
			// The Telegram API rate limit for group messages is 20 per minute. So we cannot update messages too frequently.
			minWaitDurSecs := lo.Min([]int{nUpdates, 4}) + nErrs*3
			minWait = time.After(time.Duration(minWaitDurSecs) * time.Second)

			// Send the partial message
			respText := respBuilder.String() + assistantWritingSign
			if err := cb(respText, false); err != nil {
				logger.Warnw("failed to send partial update", "error", err)
				nErrs += 1
				if nErrs > 5 {
					logger.Warnw("too many errors, aborting")
					return
				}
				continue
			}

			logger.Debugf("... message edited")
		}

		respText := respBuilder.String()
		if err = cb(respText, true); err != nil {
			logger.Warnw("assistant: failed to send message", "error", err)
		}
	}()

	return nil
}

func handleAssistantConversation(c tele.Context, thread []*tele.Message) error {
	me := c.Bot().Me
	lastMsg := thread[len(thread)-1]
	if err := cacheMessage(lastMsg); err != nil {
		logger.Warnw("failed to cache message", "error", err)
	}

	nBytes := 0                   // Used to estimated number of tokens. For now we treat 3 bytes as 1 token.
	nBytesMax := (4096 - 512) * 3 // Leave some space for the response

	sysMsg := prompts.Assistant()
	chatReqMsgs := []openai.ChatMessage{
		{
			Role:    openai.ChatRoleSystem,
			Content: sysMsg,
		},
	}
	nBytes += len(sysMsg)

	convMsgs := []openai.ChatMessage{}
	for l := len(thread) - 1; l >= 0; l-- {
		text := assistantRemoveMention(thread[l].Text, me.Username)
		textLen := len(text)
		if textLen+nBytes > nBytesMax {
			break
		}
		nBytes += textLen

		role := openai.ChatRoleUser
		from := thread[l].Sender
		if from != nil && from.ID == me.ID {
			role = openai.ChatRoleAssistant
		}

		convMsgs = append(convMsgs, openai.ChatMessage{
			Role:    role,
			Content: text,
		})
	}
	if len(convMsgs) == 0 {
		// It turns out that this will never happen because Telegram splits messages when they exceed a certain length.
		return c.Reply("Your message is too long (Sorry!)")
	}
	for l := len(convMsgs) - 1; l >= 0; l-- {
		chatReqMsgs = append(chatReqMsgs, convMsgs[l])
	}

	req := openai.ChatRequest{
		Model:       openai.ModelGpt04,
		Messages:    chatReqMsgs,
		Temperature: lo.ToPtr(0.42),
		MaxTokens:   2048,
		User:        assistantHashUserId(lastMsg.Sender.ID),
	}

	typingNotifyCh := setTyping(c)

	var replyMsg *tele.Message
	reqErr := assistantStreamedResponse(req, func(text string, finished bool) error {
		var err error
		if replyMsg == nil {
			<-typingNotifyCh
			replyMsg, err = c.Bot().Reply(lastMsg, text, tele.Silent)
		} else {
			replyMsg, err = c.Bot().Edit(replyMsg, text)
		}
		if finished && err == nil {
			replyMsg.ReplyTo = lastMsg // nasty bug
			if err := cacheMessage(replyMsg); err != nil {
				logger.Warnw("failed to cache message", "error", err)
			}
		}
		return err
	})

	if reqErr != nil {
		logger.Errorw("assistant: failed to complete conversation", "error", reqErr)
		return c.Reply("Sorry, there's a technical issue. 😵💫 Please try again later.", tele.Silent)
	}
	return nil
}

func assistantRemoveMention(msg, name string) string {
	mentionStr := "@" + name
	orig := msg
	msg = strings.TrimLeftFunc(msg, unicode.IsSpace)
	if strings.HasPrefix(msg, mentionStr) {
		return msg[len(mentionStr):]
	}
	return orig
}

func assistantHashUserId(uid int64) string {
	seasoned := []byte("RdnuRPqp66vtbc28QRO0ecKSLKXifz7G9UbXLoyCMpw" + strconv.FormatInt(uid, 10))
	hashed := md5.Sum(seasoned) // Don't judge me
	return base64.URLEncoding.EncodeToString(hashed[:])[:22]
}