feat: support channel ai.ls now (close #99)

* Make token calculation more accurate.

* fix: make the token number calculation more accurate

---------

Co-authored-by: JustSong <songquanpeng@foxmail.com>

README.md

@@ -51,9 +51,10 @@ _✨ All in one 的 OpenAI 接口，整合各种 API 访问方式，开箱即用
    + [x] **Azure OpenAI API**
    + [x] [API2D](https://api2d.com/r/197971)
    + [x] [OhMyGPT](https://aigptx.top?aff=uFpUl2Kf)
-   + [x] [CloseAI](https://console.openai-asia.com)
-   + [x] [OpenAI-SB](https://openai-sb.com)
+   + [x] [AI.LS](https://ai.ls)
    + [x] [OpenAI Max](https://openaimax.com)
+   + [x] [OpenAI-SB](https://openai-sb.com)
+   + [x] [CloseAI](https://console.openai-asia.com)
    + [x] 自定义渠道：例如使用自行搭建的 OpenAI 代理
 2. 支持通过**负载均衡**的方式访问多个渠道。
 3. 支持 **stream 模式**，可以通过流式传输实现打字机效果。
@@ -93,13 +94,10 @@ server{
           proxy_set_header X-Forwarded-For $remote_addr;
           proxy_cache_bypass $http_upgrade;
           proxy_set_header Accept-Encoding gzip;
-          proxy_buffering off;  # 重要：关闭代理缓冲
    }
 }
 ```
 
-注意，为了 SSE 正常工作，需要关闭 Nginx 的代理缓冲。
-
 之后使用 Let's Encrypt 的 certbot 配置 HTTPS：
 ```bash
 # Ubuntu 安装 certbot：

common/constants.go

@@ -127,6 +127,7 @@ const (
 	ChannelTypeOpenAIMax = 6
 	ChannelTypeOhMyGPT   = 7
 	ChannelTypeCustom    = 8
+	ChannelTypeAILS      = 9
 )
 
 var ChannelBaseURLs = []string{
@@ -139,4 +140,5 @@ var ChannelBaseURLs = []string{
 	"https://api.openaimax.com",   // 6
 	"https://api.ohmygpt.com",     // 7
 	"",                            // 8
+	"https://api.caipacity.com",   // 9
 }

controller/relay-utils.go

@@ -0,0 +1,61 @@
+package controller
+
+import (
+	"fmt"
+	"github.com/pkoukk/tiktoken-go"
+	"one-api/common"
+	"strings"
+)
+
+var tokenEncoderMap = map[string]*tiktoken.Tiktoken{}
+
+func getTokenEncoder(model string) *tiktoken.Tiktoken {
+	if tokenEncoder, ok := tokenEncoderMap[model]; ok {
+		return tokenEncoder
+	}
+	tokenEncoder, err := tiktoken.EncodingForModel(model)
+	if err != nil {
+		common.FatalLog(fmt.Sprintf("failed to get token encoder for model %s: %s", model, err.Error()))
+	}
+	tokenEncoderMap[model] = tokenEncoder
+	return tokenEncoder
+}
+
+func countTokenMessages(messages []Message, model string) int {
+	tokenEncoder := getTokenEncoder(model)
+	// Reference:
+	// https://github.com/openai/openai-cookbook/blob/main/examples/How_to_count_tokens_with_tiktoken.ipynb
+	// https://github.com/pkoukk/tiktoken-go/issues/6
+	//
+	// Every message follows <|start|>{role/name}\n{content}<|end|>\n
+	var tokensPerMessage int
+	var tokensPerName int
+	if strings.HasPrefix(model, "gpt-3.5") {
+		tokensPerMessage = 4
+		tokensPerName = -1 // If there's a name, the role is omitted
+	} else if strings.HasPrefix(model, "gpt-4") {
+		tokensPerMessage = 3
+		tokensPerName = 1
+	} else {
+		tokensPerMessage = 3
+		tokensPerName = 1
+	}
+	tokenNum := 0
+	for _, message := range messages {
+		tokenNum += tokensPerMessage
+		tokenNum += len(tokenEncoder.Encode(message.Content, nil, nil))
+		tokenNum += len(tokenEncoder.Encode(message.Role, nil, nil))
+		if message.Name != nil {
+			tokenNum += tokensPerName
+			tokenNum += len(tokenEncoder.Encode(*message.Name, nil, nil))
+		}
+	}
+	tokenNum += 3 // Every reply is primed with <|start|>assistant<|message|>
+	return tokenNum
+}
+
+func countTokenText(text string, model string) int {
+	tokenEncoder := getTokenEncoder(model)
+	token := tokenEncoder.Encode(text, nil, nil)
+	return len(token)
+}

controller/relay.go

@@ -6,7 +6,6 @@ import (
 	"encoding/json"
 	"fmt"
 	"github.com/gin-gonic/gin"
-	"github.com/pkoukk/tiktoken-go"
 	"io"
 	"net/http"
 	"one-api/common"
@@ -15,8 +14,9 @@ import (
 )
 
 type Message struct {
-	Role    string `json:"role"`
-	Content string `json:"content"`
+	Role    string  `json:"role"`
+	Content string  `json:"content"`
+	Name    *string `json:"name,omitempty"`
 }
 
 type ChatRequest struct {
@@ -65,13 +65,6 @@ type StreamResponse struct {
 	} `json:"choices"`
 }
 
-var tokenEncoder, _ = tiktoken.GetEncoding("cl100k_base")
-
-func countToken(text string) int {
-	token := tokenEncoder.Encode(text, nil, nil)
-	return len(token)
-}
-
 func Relay(c *gin.Context) {
 	err := relayHelper(c)
 	if err != nil {
@@ -149,11 +142,8 @@ func relayHelper(c *gin.Context) *OpenAIErrorWithStatusCode {
 		model_ = strings.TrimSuffix(model_, "-0314")
 		fullRequestURL = fmt.Sprintf("%s/openai/deployments/%s/%s", baseURL, model_, task)
 	}
-	var promptText string
-	for _, message := range textRequest.Messages {
-		promptText += fmt.Sprintf("%s: %s\n", message.Role, message.Content)
-	}
-	promptTokens := countToken(promptText) + 3
+
+	promptTokens := countTokenMessages(textRequest.Messages, textRequest.Model)
 	preConsumedTokens := common.PreConsumedQuota
 	if textRequest.MaxTokens != 0 {
 		preConsumedTokens = promptTokens + textRequest.MaxTokens
@@ -206,8 +196,8 @@ func relayHelper(c *gin.Context) *OpenAIErrorWithStatusCode {
 				completionRatio = 2
 			}
 			if isStream {
-				completionText := fmt.Sprintf("%s: %s\n", "assistant", streamResponseText)
-				quota = promptTokens + countToken(completionText)*completionRatio
+				responseTokens := countTokenText(streamResponseText, textRequest.Model)
+				quota = promptTokens + responseTokens*completionRatio
 			} else {
 				quota = textResponse.Usage.PromptTokens + textResponse.Usage.CompletionTokens*completionRatio
 			}
@@ -242,6 +232,10 @@ func relayHelper(c *gin.Context) *OpenAIErrorWithStatusCode {
 		go func() {
 			for scanner.Scan() {
 				data := scanner.Text()
+				if len(data) < 6 { // must be something wrong!
+					common.SysError("Invalid stream response: " + data)
+					continue
+				}
 				dataChan <- data
 				data = data[6:]
 				if !strings.HasPrefix(data, "[DONE]") {

web/src/constants/channel.constants.js

@@ -6,5 +6,6 @@ export const CHANNEL_OPTIONS = [
   { key: 5, text: 'OpenAI-SB', value: 5, color: 'brown' },
   { key: 6, text: 'OpenAI Max', value: 6, color: 'violet' },
   { key: 7, text: 'OhMyGPT', value: 7, color: 'purple' },
+  { key: 9, text: 'AI.LS', value: 9, color: 'yellow' },
   { key: 8, text: '自定义', value: 8, color: 'pink' }
 ];