feat: support gpt-4 with vision (#683, #714)

* fix: Skip JSON deserialization when accessing transcriptions and translations.

* chore: update impl

---------

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

README.md

@@ -92,14 +92,14 @@ _✨ 通过标准的 OpenAI API 格式访问所有的大模型，开箱即用
 12. 支持**用户邀请奖励**。
 13. 支持以美元为单位显示额度。
 14. 支持发布公告，设置充值链接，设置新用户初始额度。
-15. 支持模型映射，重定向用户的请求模型。
+15. 支持模型映射，重定向用户的请求模型，如无必要请不要设置，设置之后会导致请求体被重新构造而非直接透传，会导致部分还未正式支持的字段无法传递成功。
 16. 支持失败自动重试。
 17. 支持绘图接口。
 18. 支持 [Cloudflare AI Gateway](https://developers.cloudflare.com/ai-gateway/providers/openai/)，渠道设置的代理部分填写 `https://gateway.ai.cloudflare.com/v1/ACCOUNT_TAG/GATEWAY/openai` 即可。
 19. 支持丰富的**自定义**设置，
     1. 支持自定义系统名称，logo 以及页脚。
     2. 支持自定义首页和关于页面，可以选择使用 HTML & Markdown 代码进行自定义，或者使用一个单独的网页通过 iframe 嵌入。
-20. 支持通过系统访问令牌访问管理 API。
+20. 支持通过系统访问令牌访问管理 API（bearer token，用以替代 cookie，你可以自行抓包来查看 API 的用法）。
 21. 支持 Cloudflare Turnstile 用户校验。
 22. 支持用户管理，支持**多种用户登录注册方式**：
     + 邮箱登录注册（支持注册邮箱白名单）以及通过邮箱进行密码重置。

common/gin.go

@@ -5,6 +5,7 @@ import (
 	"encoding/json"
 	"github.com/gin-gonic/gin"
 	"io"
+	"strings"
 )
 
 func UnmarshalBodyReusable(c *gin.Context, v any) error {
@@ -16,7 +17,13 @@ func UnmarshalBodyReusable(c *gin.Context, v any) error {
 	if err != nil {
 		return err
 	}
-	err = json.Unmarshal(requestBody, &v)
+	contentType := c.Request.Header.Get("Content-Type")
+	if strings.HasPrefix(contentType, "application/json") {
+		err = json.Unmarshal(requestBody, &v)
+	} else {
+		// skip for now
+		// TODO: someday non json request have variant model, we will need to implementation this
+	}
 	if err != nil {
 		return err
 	}

controller/channel-test.go

@@ -5,14 +5,15 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
-	"github.com/gin-gonic/gin"
+	"io"
 	"net/http"
 	"one-api/common"
 	"one-api/model"
 	"strconv"
-	"strings"
 	"sync"
 	"time"
+
+	"github.com/gin-gonic/gin"
 )
 
 func testChannel(channel *model.Channel, request ChatRequest) (err error, openaiErr *OpenAIError) {
@@ -43,16 +44,14 @@ func testChannel(channel *model.Channel, request ChatRequest) (err error, openai
 	}
 	requestURL := common.ChannelBaseURLs[channel.Type]
 	if channel.Type == common.ChannelTypeAzure {
-		requestURL = fmt.Sprintf("%s/openai/deployments/%s/chat/completions?api-version=2023-03-15-preview", channel.GetBaseURL(), request.Model)
+		requestURL = getFullRequestURL(channel.GetBaseURL(), fmt.Sprintf("/openai/deployments/%s/chat/completions?api-version=2023-03-15-preview", request.Model), channel.Type)
 	} else {
-		if channel.GetBaseURL() != "" {
-			requestURL = channel.GetBaseURL()
+		if baseURL := channel.GetBaseURL(); len(baseURL) > 0 {
+			requestURL = baseURL
 		}
-		requestURL += "/v1/chat/completions"
-	}
-	// for Cloudflare AI gateway: https://github.com/songquanpeng/one-api/pull/639
-	requestURL = strings.Replace(requestURL, "/v1/v1", "/v1", 1)
 
+		requestURL = getFullRequestURL(requestURL, "/v1/chat/completions", channel.Type)
+	}
 	jsonData, err := json.Marshal(request)
 	if err != nil {
 		return err, nil
@@ -73,10 +72,14 @@ func testChannel(channel *model.Channel, request ChatRequest) (err error, openai
 	}
 	defer resp.Body.Close()
 	var response TextResponse
-	err = json.NewDecoder(resp.Body).Decode(&response)
+	body, err := io.ReadAll(resp.Body)
 	if err != nil {
 		return err, nil
 	}
+	err = json.Unmarshal(body, &response)
+	if err != nil {
+		return fmt.Errorf("Error: %s\nResp body: %s", err, body), nil
+	}
 	if response.Usage.CompletionTokens == 0 {
 		return errors.New(fmt.Sprintf("type %s, code %v, message %s", response.Error.Type, response.Error.Code, response.Error.Message)), &response.Error
 	}

controller/relay-aiproxy.go

@@ -48,7 +48,7 @@ type AIProxyLibraryStreamResponse struct {
 func requestOpenAI2AIProxyLibrary(request GeneralOpenAIRequest) *AIProxyLibraryRequest {
 	query := ""
 	if len(request.Messages) != 0 {
-		query = request.Messages[len(request.Messages)-1].Content
+		query = request.Messages[len(request.Messages)-1].StringContent()
 	}
 	return &AIProxyLibraryRequest{
 		Model:  request.Model,

controller/relay-ali.go

@@ -88,18 +88,18 @@ func requestOpenAI2Ali(request GeneralOpenAIRequest) *AliChatRequest {
 		message := request.Messages[i]
 		if message.Role == "system" {
 			messages = append(messages, AliMessage{
-				User: message.Content,
+				User: message.StringContent(),
 				Bot:  "Okay",
 			})
 			continue
 		} else {
 			if i == len(request.Messages)-1 {
-				prompt = message.Content
+				prompt = message.StringContent()
 				break
 			}
 			messages = append(messages, AliMessage{
-				User: message.Content,
-				Bot:  request.Messages[i+1].Content,
+				User: message.StringContent(),
+				Bot:  request.Messages[i+1].StringContent(),
 			})
 			i++
 		}

controller/relay-baidu.go

@@ -89,7 +89,7 @@ func requestOpenAI2Baidu(request GeneralOpenAIRequest) *BaiduChatRequest {
 		if message.Role == "system" {
 			messages = append(messages, BaiduMessage{
 				Role:    "user",
-				Content: message.Content,
+				Content: message.StringContent(),
 			})
 			messages = append(messages, BaiduMessage{
 				Role:    "assistant",
@@ -98,7 +98,7 @@ func requestOpenAI2Baidu(request GeneralOpenAIRequest) *BaiduChatRequest {
 		} else {
 			messages = append(messages, BaiduMessage{
 				Role:    message.Role,
-				Content: message.Content,
+				Content: message.StringContent(),
 			})
 		}
 	}

controller/relay-openai.go

@@ -132,7 +132,7 @@ func openaiHandler(c *gin.Context, resp *http.Response, consumeQuota bool, promp
 	if textResponse.Usage.TotalTokens == 0 {
 		completionTokens := 0
 		for _, choice := range textResponse.Choices {
-			completionTokens += countTokenText(choice.Message.Content, model)
+			completionTokens += countTokenText(choice.Message.StringContent(), model)
 		}
 		textResponse.Usage = Usage{
 			PromptTokens:     promptTokens,

controller/relay-palm.go

@@ -59,7 +59,7 @@ func requestOpenAI2PaLM(textRequest GeneralOpenAIRequest) *PaLMChatRequest {
 	}
 	for _, message := range textRequest.Messages {
 		palmMessage := PaLMChatMessage{
-			Content: message.Content,
+			Content: message.StringContent(),
 		}
 		if message.Role == "user" {
 			palmMessage.Author = "0"

controller/relay-tencent.go

@@ -84,7 +84,7 @@ func requestOpenAI2Tencent(request GeneralOpenAIRequest) *TencentChatRequest {
 		if message.Role == "system" {
 			messages = append(messages, TencentMessage{
 				Role:    "user",
-				Content: message.Content,
+				Content: message.StringContent(),
 			})
 			messages = append(messages, TencentMessage{
 				Role:    "assistant",
@@ -93,7 +93,7 @@ func requestOpenAI2Tencent(request GeneralOpenAIRequest) *TencentChatRequest {
 			continue
 		}
 		messages = append(messages, TencentMessage{
-			Content: message.Content,
+			Content: message.StringContent(),
 			Role:    message.Role,
 		})
 	}

controller/relay-text.go

@@ -147,7 +147,9 @@ func relayTextHelper(c *gin.Context, relayMode int) *OpenAIErrorWithStatusCode {
 			model_ = strings.TrimSuffix(model_, "-0301")
 			model_ = strings.TrimSuffix(model_, "-0314")
 			model_ = strings.TrimSuffix(model_, "-0613")
-			fullRequestURL = fmt.Sprintf("%s/openai/deployments/%s/%s", baseURL, model_, task)
+
+			requestURL = fmt.Sprintf("/openai/deployments/%s/%s", model_, task)
+			fullRequestURL = getFullRequestURL(baseURL, requestURL, channelType)
 		}
 	case APITypeClaude:
 		fullRequestURL = "https://api.anthropic.com/v1/complete"

controller/relay-utils.go

@@ -4,14 +4,15 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
-	"github.com/gin-gonic/gin"
-	"github.com/pkoukk/tiktoken-go"
 	"io"
 	"net/http"
 	"one-api/common"
 	"one-api/model"
 	"strconv"
 	"strings"
+
+	"github.com/gin-gonic/gin"
+	"github.com/pkoukk/tiktoken-go"
 )
 
 var stopFinishReason = "stop"
@@ -86,7 +87,7 @@ func countTokenMessages(messages []Message, model string) int {
 	tokenNum := 0
 	for _, message := range messages {
 		tokenNum += tokensPerMessage
-		tokenNum += getTokenNum(tokenEncoder, message.Content)
+		tokenNum += getTokenNum(tokenEncoder, message.StringContent())
 		tokenNum += getTokenNum(tokenEncoder, message.Role)
 		if message.Name != nil {
 			tokenNum += tokensPerName
@@ -181,11 +182,16 @@ func relayErrorHandler(resp *http.Response) (openAIErrorWithStatusCode *OpenAIEr
 
 func getFullRequestURL(baseURL string, requestURL string, channelType int) string {
 	fullRequestURL := fmt.Sprintf("%s%s", baseURL, requestURL)
-	if channelType == common.ChannelTypeOpenAI {
-		if strings.HasPrefix(baseURL, "https://gateway.ai.cloudflare.com") {
+
+	if strings.HasPrefix(baseURL, "https://gateway.ai.cloudflare.com") {
+		switch channelType {
+		case common.ChannelTypeOpenAI:
 			fullRequestURL = fmt.Sprintf("%s%s", baseURL, strings.TrimPrefix(requestURL, "/v1"))
+		case common.ChannelTypeAzure:
+			fullRequestURL = fmt.Sprintf("%s%s", baseURL, strings.TrimPrefix(requestURL, "/openai/deployments"))
 		}
 	}
+
 	return fullRequestURL
 }
 

controller/relay-xunfei.go

@@ -81,7 +81,7 @@ func requestOpenAI2Xunfei(request GeneralOpenAIRequest, xunfeiAppId string, doma
 		if message.Role == "system" {
 			messages = append(messages, XunfeiMessage{
 				Role:    "user",
-				Content: message.Content,
+				Content: message.StringContent(),
 			})
 			messages = append(messages, XunfeiMessage{
 				Role:    "assistant",
@@ -90,7 +90,7 @@ func requestOpenAI2Xunfei(request GeneralOpenAIRequest, xunfeiAppId string, doma
 		} else {
 			messages = append(messages, XunfeiMessage{
 				Role:    message.Role,
-				Content: message.Content,
+				Content: message.StringContent(),
 			})
 		}
 	}

controller/relay-zhipu.go

@@ -114,7 +114,7 @@ func requestOpenAI2Zhipu(request GeneralOpenAIRequest) *ZhipuRequest {
 		if message.Role == "system" {
 			messages = append(messages, ZhipuMessage{
 				Role:    "system",
-				Content: message.Content,
+				Content: message.StringContent(),
 			})
 			messages = append(messages, ZhipuMessage{
 				Role:    "user",
@@ -123,7 +123,7 @@ func requestOpenAI2Zhipu(request GeneralOpenAIRequest) *ZhipuRequest {
 		} else {
 			messages = append(messages, ZhipuMessage{
 				Role:    message.Role,
-				Content: message.Content,
+				Content: message.StringContent(),
 			})
 		}
 	}

controller/relay.go

@@ -12,10 +12,49 @@ import (
 
 type Message struct {
 	Role    string  `json:"role"`
-	Content string  `json:"content"`
+	Content any     `json:"content"`
 	Name    *string `json:"name,omitempty"`
 }
 
+type ImageURL struct {
+	Url    string `json:"url,omitempty"`
+	Detail string `json:"detail,omitempty"`
+}
+
+type TextContent struct {
+	Type string `json:"type,omitempty"`
+	Text string `json:"text,omitempty"`
+}
+
+type ImageContent struct {
+	Type     string    `json:"type,omitempty"`
+	ImageURL *ImageURL `json:"image_url,omitempty"`
+}
+
+func (m Message) StringContent() string {
+	content, ok := m.Content.(string)
+	if ok {
+		return content
+	}
+	contentList, ok := m.Content.([]any)
+	if ok {
+		var contentStr string
+		for _, contentItem := range contentList {
+			contentMap, ok := contentItem.(map[string]any)
+			if !ok {
+				continue
+			}
+			if contentMap["type"] == "text" {
+				if subStr, ok := contentMap["text"].(string); ok {
+					contentStr += subStr
+				}
+			}
+		}
+		return contentStr
+	}
+	return ""
+}
+
 const (
 	RelayModeUnknown = iota
 	RelayModeChatCompletions
@@ -31,19 +70,30 @@ const (
 
 // https://platform.openai.com/docs/api-reference/chat
 
+type ResponseFormat struct {
+	Type string `json:"type,omitempty"`
+}
+
 type GeneralOpenAIRequest struct {
-	Model       string    `json:"model,omitempty"`
-	Messages    []Message `json:"messages,omitempty"`
-	Prompt      any       `json:"prompt,omitempty"`
-	Stream      bool      `json:"stream,omitempty"`
-	MaxTokens   int       `json:"max_tokens,omitempty"`
-	Temperature float64   `json:"temperature,omitempty"`
-	TopP        float64   `json:"top_p,omitempty"`
-	N           int       `json:"n,omitempty"`
-	Input       any       `json:"input,omitempty"`
-	Instruction string    `json:"instruction,omitempty"`
-	Size        string    `json:"size,omitempty"`
-	Functions   any       `json:"functions,omitempty"`
+	Model            string          `json:"model,omitempty"`
+	Messages         []Message       `json:"messages,omitempty"`
+	Prompt           any             `json:"prompt,omitempty"`
+	Stream           bool            `json:"stream,omitempty"`
+	MaxTokens        int             `json:"max_tokens,omitempty"`
+	Temperature      float64         `json:"temperature,omitempty"`
+	TopP             float64         `json:"top_p,omitempty"`
+	N                int             `json:"n,omitempty"`
+	Input            any             `json:"input,omitempty"`
+	Instruction      string          `json:"instruction,omitempty"`
+	Size             string          `json:"size,omitempty"`
+	Functions        any             `json:"functions,omitempty"`
+	FrequencyPenalty float64         `json:"frequency_penalty,omitempty"`
+	PresencePenalty  float64         `json:"presence_penalty,omitempty"`
+	ResponseFormat   *ResponseFormat `json:"response_format,omitempty"`
+	Seed             float64         `json:"seed,omitempty"`
+	Tools            any             `json:"tools,omitempty"`
+	ToolChoice       any             `json:"tool_choice,omitempty"`
+	User             string          `json:"user,omitempty"`
 }
 
 func (r GeneralOpenAIRequest) ParseInput() []string {
@@ -201,9 +251,9 @@ func Relay(c *gin.Context) {
 		relayMode = RelayModeEdits
 	} else if strings.HasPrefix(c.Request.URL.Path, "/v1/audio/speech") {
 		relayMode = RelayModeAudioSpeech
-	} else if strings.HasPrefix(c.Request.URL.Path, "/v1/audio/transcription") {
+	} else if strings.HasPrefix(c.Request.URL.Path, "/v1/audio/transcriptions") {
 		relayMode = RelayModeAudioTranscription
-	} else if strings.HasPrefix(c.Request.URL.Path, "/v1/audio/translation") {
+	} else if strings.HasPrefix(c.Request.URL.Path, "/v1/audio/translations") {
 		relayMode = RelayModeAudioTranslation
 	}
 	var err *OpenAIErrorWithStatusCode