fix: use gpt-3.5's encoder if not found (close #110)

* Make token calculation more accurate.

* fix: make the token number calculation more accurate

---------

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

README.md

@@ -51,15 +51,17 @@ _✨ 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 Proxy](https://aiproxy.io/?i=OneAPI) （邀请码：`OneAPI`）
+   + [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/r/2412)
    + [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。
 4. 支持**多机部署**，[详见此处](#多机部署)。
 5. 支持**令牌管理**，设置令牌的过期时间和使用次数。
-6. 支持**兑换码管理**，支持批量生成和导出兑换码，可使用兑换码为令牌进行充值。
+6. 支持**兑换码管理**，支持批量生成和导出兑换码，可使用兑换码为账户进行充值。
 7. 支持**通道管理**，批量创建通道。
 8. 支持发布公告，设置充值链接，设置新用户初始额度。
 9. 支持丰富的**自定义**设置，
@@ -93,13 +95,10 @@ 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 +127,8 @@ const (
 	ChannelTypeOpenAIMax = 6
 	ChannelTypeOhMyGPT   = 7
 	ChannelTypeCustom    = 8
+	ChannelTypeAILS      = 9
+	ChannelTypeAIProxy   = 10
 )
 
 var ChannelBaseURLs = []string{
@@ -139,4 +141,6 @@ var ChannelBaseURLs = []string{
 	"https://api.openaimax.com",   // 6
 	"https://api.ohmygpt.com",     // 7
 	"",                            // 8
+	"https://api.caipacity.com",   // 9
+	"https://api.aiproxy.io",      // 10
 }

controller/relay-utils.go

@@ -0,0 +1,65 @@
+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.SysError(fmt.Sprintf("failed to get token encoder for model %s: %s, using encoder for gpt-3.5-turbo", model, err.Error()))
+		tokenEncoder, err = tiktoken.EncodingForModel("gpt-3.5-turbo")
+		if err != nil {
+			common.FatalLog(fmt.Sprintf("failed to get token encoder for model gpt-3.5-turbo: %s", 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]") {

model/main.go

@@ -26,6 +26,7 @@ func createRootAccountIfNeed() error {
 			Status:      common.UserStatusEnabled,
 			DisplayName: "Root User",
 			AccessToken: common.GetUUID(),
+			Quota:       100000000,
 		}
 		DB.Create(&rootUser)
 	}

model/user.go

@@ -19,8 +19,7 @@ type User struct {
 	Email            string `json:"email" gorm:"index" validate:"max=50"`
 	GitHubId         string `json:"github_id" gorm:"column:github_id;index"`
 	WeChatId         string `json:"wechat_id" gorm:"column:wechat_id;index"`
-	VerificationCode string `json:"verification_code" gorm:"-:all"` // this field is only for Email verification, don't save it to database!
-	Balance          int    `json:"balance" gorm:"type:int;default:0"`
+	VerificationCode string `json:"verification_code" gorm:"-:all"`                                    // this field is only for Email verification, don't save it to database!
 	AccessToken      string `json:"access_token" gorm:"type:char(32);column:access_token;uniqueIndex"` // this token is for system management
 	Quota            int    `json:"quota" gorm:"type:int;default:0"`
 }

web/src/constants/channel.constants.js

@@ -1,10 +1,12 @@
 export const CHANNEL_OPTIONS = [
   { key: 1, text: 'OpenAI', value: 1, color: 'green' },
-  { key: 2, text: 'API2D', value: 2, color: 'blue' },
+  { key: 8, text: '自定义', value: 8, color: 'pink' },
   { key: 3, text: 'Azure', value: 3, color: 'olive' },
+  { key: 2, text: 'API2D', value: 2, color: 'blue' },
   { key: 4, text: 'CloseAI', value: 4, color: 'teal' },
   { 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: 8, text: '自定义', value: 8, color: 'pink' }
+  { key: 9, text: 'AI.LS', value: 9, color: 'yellow' },
+  { key: 10, text: 'AI Proxy', value: 10, color: 'purple' }
 ];

web/src/pages/Channel/EditChannel.js

@@ -46,6 +46,9 @@ const EditChannel = () => {
     if (localInputs.base_url.endsWith('/')) {
       localInputs.base_url = localInputs.base_url.slice(0, localInputs.base_url.length - 1);
     }
+    if (localInputs.type === 3 && localInputs.other === '') {
+      localInputs.other = '2023-03-15-preview';
+    }
     let res;
     if (isEdit) {
       res = await API.put(`/api/channel/`, { ...localInputs, id: parseInt(channelId) });

web/src/pages/TopUp/index.js

@@ -1,6 +1,6 @@
 import React, { useEffect, useState } from 'react';
 import { Button, Form, Grid, Header, Segment, Statistic } from 'semantic-ui-react';
-import { API, showError, showSuccess } from '../../helpers';
+import { API, showError, showInfo, showSuccess } from '../../helpers';
 
 const TopUp = () => {
   const [redemptionCode, setRedemptionCode] = useState('');
@@ -9,6 +9,7 @@ const TopUp = () => {
 
   const topUp = async () => {
     if (redemptionCode === '') {
+      showInfo('请输入充值码！')
       return;
     }
     const res = await API.post('/api/user/topup', {
@@ -80,7 +81,7 @@ const TopUp = () => {
         <Grid.Column>
           <Statistic.Group widths='one'>
             <Statistic>
-              <Statistic.Value>{userQuota}</Statistic.Value>
+              <Statistic.Value>{userQuota.toLocaleString()}</Statistic.Value>
               <Statistic.Label>剩余额度</Statistic.Label>
             </Statistic>
           </Statistic.Group>

web/src/pages/User/EditUser.js

@@ -14,8 +14,9 @@ const EditUser = () => {
     github_id: '',
     wechat_id: '',
     email: '',
+    quota: 0,
   });
-  const { username, display_name, password, github_id, wechat_id, email } =
+  const { username, display_name, password, github_id, wechat_id, email, quota } =
     inputs;
   const handleInputChange = (e, { name, value }) => {
     setInputs((inputs) => ({ ...inputs, [name]: value }));
@@ -44,7 +45,11 @@ const EditUser = () => {
   const submit = async () => {
     let res = undefined;
     if (userId) {
-      res = await API.put(`/api/user/`, { ...inputs, id: parseInt(userId) });
+      let data = { ...inputs, id: parseInt(userId) };
+      if (typeof data.quota === 'string') {
+        data.quota = parseInt(data.quota);
+      }
+      res = await API.put(`/api/user/`, data);
     } else {
       res = await API.put(`/api/user/self`, inputs);
     }
@@ -92,6 +97,21 @@ const EditUser = () => {
               autoComplete='new-password'
             />
           </Form.Field>
+          {
+            userId && (
+              <Form.Field>
+                <Form.Input
+                  label='剩余额度'
+                  name='quota'
+                  placeholder={'请输入新的剩余额度'}
+                  onChange={handleInputChange}
+                  value={quota}
+                  type={'number'}
+                  autoComplete='new-password'
+                />
+              </Form.Field>
+            )
+          }
           <Form.Field>
             <Form.Input
               label='已绑定的 GitHub 账户'