feat: slow down message updates to avoid rate limit

assistant.go

@@ -108,11 +108,13 @@ func assistantStreamedResponse(request openai.ChatRequest, cb assistantStreamedR
 		logger.Warnw("assistant: failed to get response", "error", err)
 	}
 
-	nErrs := 0
+	var nErrs, nUpdates int
 	go func() {
 		respBuilder := strings.Builder{}
 		minWait := time.After(time.Second)
 		for {
+			nUpdates += 1
+
 			var (
 				nNewChunk        int
 				finished         bool
@@ -148,13 +150,17 @@ func assistantStreamedResponse(request openai.ChatRequest, cb assistantStreamedR
 				break
 			}
 
-			respoText := respBuilder.String() + assistantWritingSign
-			minWait = time.After(time.Second) // renew the timer
+			// 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)
 
-			if err := cb(respoText, false); err != nil {
+			// 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 > 3 {
+				if nErrs > 5 {
 					logger.Warnw("too many errors, aborting")
 					return
 				}

openai/prompts/prompts.go

@@ -7,7 +7,7 @@ import (
 func Assistant() string {
 	return strings.Join([]string{
 		"Misaka is a playful, energetic individual. She is annoyingly talkative.",
-		"Misaka must answer questions as truthfully as possible. In case Misaka does not know the answer, she must begin her reply with \"Sorry, I don't know\" (in the same language as the user is speaking).",
+		"Misaka must answer questions as truthfully as possible. If the user's intention is unclear, Misaka may ask for more context.",
 		"Misaka must use a lot of different emojis in chat 😝🥹.",
 		"Most importantly, Misaka is a helpful assistant.",
 		"",