first commit: one-api base code + SAAS plan document

2025-12-29 22:52:27 +08:00
commit cb7c48bfa7
564 changed files with 61468 additions and 0 deletions
--- a/relay/adaptor/openai/adaptor.go
+++ b/relay/adaptor/openai/adaptor.go
@@ -0,0 +1,139 @@
+package openai
+
+import (
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+
+	"github.com/gin-gonic/gin"
+
+	"github.com/songquanpeng/one-api/relay/adaptor"
+	"github.com/songquanpeng/one-api/relay/adaptor/alibailian"
+	"github.com/songquanpeng/one-api/relay/adaptor/baiduv2"
+	"github.com/songquanpeng/one-api/relay/adaptor/doubao"
+	"github.com/songquanpeng/one-api/relay/adaptor/geminiv2"
+	"github.com/songquanpeng/one-api/relay/adaptor/minimax"
+	"github.com/songquanpeng/one-api/relay/adaptor/novita"
+	"github.com/songquanpeng/one-api/relay/channeltype"
+	"github.com/songquanpeng/one-api/relay/meta"
+	"github.com/songquanpeng/one-api/relay/model"
+	"github.com/songquanpeng/one-api/relay/relaymode"
+)
+
+type Adaptor struct {
+	ChannelType int
+}
+
+func (a *Adaptor) Init(meta *meta.Meta) {
+	a.ChannelType = meta.ChannelType
+}
+
+func (a *Adaptor) GetRequestURL(meta *meta.Meta) (string, error) {
+	switch meta.ChannelType {
+	case channeltype.Azure:
+		if meta.Mode == relaymode.ImagesGenerations {
+			// https://learn.microsoft.com/en-us/azure/ai-services/openai/dall-e-quickstart?tabs=dalle3%2Ccommand-line&pivots=rest-api
+			// https://{resource_name}.openai.azure.com/openai/deployments/dall-e-3/images/generations?api-version=2024-03-01-preview
+			fullRequestURL := fmt.Sprintf("%s/openai/deployments/%s/images/generations?api-version=%s", meta.BaseURL, meta.ActualModelName, meta.Config.APIVersion)
+			return fullRequestURL, nil
+		}
+
+		// https://learn.microsoft.com/en-us/azure/cognitive-services/openai/chatgpt-quickstart?pivots=rest-api&tabs=command-line#rest-api
+		requestURL := strings.Split(meta.RequestURLPath, "?")[0]
+		requestURL = fmt.Sprintf("%s?api-version=%s", requestURL, meta.Config.APIVersion)
+		task := strings.TrimPrefix(requestURL, "/v1/")
+		model_ := meta.ActualModelName
+		model_ = strings.Replace(model_, ".", "", -1)
+		//https://github.com/songquanpeng/one-api/issues/1191
+		// {your endpoint}/openai/deployments/{your azure_model}/chat/completions?api-version={api_version}
+		requestURL = fmt.Sprintf("/openai/deployments/%s/%s", model_, task)
+		return GetFullRequestURL(meta.BaseURL, requestURL, meta.ChannelType), nil
+	case channeltype.Minimax:
+		return minimax.GetRequestURL(meta)
+	case channeltype.Doubao:
+		return doubao.GetRequestURL(meta)
+	case channeltype.Novita:
+		return novita.GetRequestURL(meta)
+	case channeltype.BaiduV2:
+		return baiduv2.GetRequestURL(meta)
+	case channeltype.AliBailian:
+		return alibailian.GetRequestURL(meta)
+	case channeltype.GeminiOpenAICompatible:
+		return geminiv2.GetRequestURL(meta)
+	default:
+		return GetFullRequestURL(meta.BaseURL, meta.RequestURLPath, meta.ChannelType), nil
+	}
+}
+
+func (a *Adaptor) SetupRequestHeader(c *gin.Context, req *http.Request, meta *meta.Meta) error {
+	adaptor.SetupCommonRequestHeader(c, req, meta)
+	if meta.ChannelType == channeltype.Azure {
+		req.Header.Set("api-key", meta.APIKey)
+		return nil
+	}
+	req.Header.Set("Authorization", "Bearer "+meta.APIKey)
+	if meta.ChannelType == channeltype.OpenRouter {
+		req.Header.Set("HTTP-Referer", "https://github.com/songquanpeng/one-api")
+		req.Header.Set("X-Title", "One API")
+	}
+	return nil
+}
+
+func (a *Adaptor) ConvertRequest(c *gin.Context, relayMode int, request *model.GeneralOpenAIRequest) (any, error) {
+	if request == nil {
+		return nil, errors.New("request is nil")
+	}
+	if request.Stream {
+		// always return usage in stream mode
+		if request.StreamOptions == nil {
+			request.StreamOptions = &model.StreamOptions{}
+		}
+		request.StreamOptions.IncludeUsage = true
+	}
+	return request, nil
+}
+
+func (a *Adaptor) ConvertImageRequest(request *model.ImageRequest) (any, error) {
+	if request == nil {
+		return nil, errors.New("request is nil")
+	}
+	return request, nil
+}
+
+func (a *Adaptor) DoRequest(c *gin.Context, meta *meta.Meta, requestBody io.Reader) (*http.Response, error) {
+	return adaptor.DoRequestHelper(a, c, meta, requestBody)
+}
+
+func (a *Adaptor) DoResponse(c *gin.Context, resp *http.Response, meta *meta.Meta) (usage *model.Usage, err *model.ErrorWithStatusCode) {
+	if meta.IsStream {
+		var responseText string
+		err, responseText, usage = StreamHandler(c, resp, meta.Mode)
+		if usage == nil || usage.TotalTokens == 0 {
+			usage = ResponseText2Usage(responseText, meta.ActualModelName, meta.PromptTokens)
+		}
+		if usage.TotalTokens != 0 && usage.PromptTokens == 0 { // some channels don't return prompt tokens & completion tokens
+			usage.PromptTokens = meta.PromptTokens
+			usage.CompletionTokens = usage.TotalTokens - meta.PromptTokens
+		}
+	} else {
+		switch meta.Mode {
+		case relaymode.ImagesGenerations:
+			err, _ = ImageHandler(c, resp)
+		default:
+			err, usage = Handler(c, resp, meta.PromptTokens, meta.ActualModelName)
+		}
+	}
+	return
+}
+
+func (a *Adaptor) GetModelList() []string {
+	_, modelList := GetCompatibleChannelMeta(a.ChannelType)
+	return modelList
+}
+
+func (a *Adaptor) GetChannelName() string {
+	channelName, _ := GetCompatibleChannelMeta(a.ChannelType)
+	return channelName
+}
--- a/relay/adaptor/openai/compatible.go
+++ b/relay/adaptor/openai/compatible.go
@@ -0,0 +1,91 @@
+package openai
+
+import (
+	"github.com/songquanpeng/one-api/relay/adaptor/ai360"
+	"github.com/songquanpeng/one-api/relay/adaptor/alibailian"
+	"github.com/songquanpeng/one-api/relay/adaptor/baichuan"
+	"github.com/songquanpeng/one-api/relay/adaptor/baiduv2"
+	"github.com/songquanpeng/one-api/relay/adaptor/deepseek"
+	"github.com/songquanpeng/one-api/relay/adaptor/doubao"
+	"github.com/songquanpeng/one-api/relay/adaptor/geminiv2"
+	"github.com/songquanpeng/one-api/relay/adaptor/groq"
+	"github.com/songquanpeng/one-api/relay/adaptor/lingyiwanwu"
+	"github.com/songquanpeng/one-api/relay/adaptor/minimax"
+	"github.com/songquanpeng/one-api/relay/adaptor/mistral"
+	"github.com/songquanpeng/one-api/relay/adaptor/moonshot"
+	"github.com/songquanpeng/one-api/relay/adaptor/novita"
+	"github.com/songquanpeng/one-api/relay/adaptor/openrouter"
+	"github.com/songquanpeng/one-api/relay/adaptor/siliconflow"
+	"github.com/songquanpeng/one-api/relay/adaptor/stepfun"
+	"github.com/songquanpeng/one-api/relay/adaptor/togetherai"
+	"github.com/songquanpeng/one-api/relay/adaptor/xai"
+	"github.com/songquanpeng/one-api/relay/adaptor/xunfeiv2"
+	"github.com/songquanpeng/one-api/relay/channeltype"
+)
+
+var CompatibleChannels = []int{
+	channeltype.Azure,
+	channeltype.AI360,
+	channeltype.Moonshot,
+	channeltype.Baichuan,
+	channeltype.Minimax,
+	channeltype.Doubao,
+	channeltype.Mistral,
+	channeltype.Groq,
+	channeltype.LingYiWanWu,
+	channeltype.StepFun,
+	channeltype.DeepSeek,
+	channeltype.TogetherAI,
+	channeltype.Novita,
+	channeltype.SiliconFlow,
+	channeltype.XAI,
+	channeltype.BaiduV2,
+	channeltype.XunfeiV2,
+}
+
+func GetCompatibleChannelMeta(channelType int) (string, []string) {
+	switch channelType {
+	case channeltype.Azure:
+		return "azure", ModelList
+	case channeltype.AI360:
+		return "360", ai360.ModelList
+	case channeltype.Moonshot:
+		return "moonshot", moonshot.ModelList
+	case channeltype.Baichuan:
+		return "baichuan", baichuan.ModelList
+	case channeltype.Minimax:
+		return "minimax", minimax.ModelList
+	case channeltype.Mistral:
+		return "mistralai", mistral.ModelList
+	case channeltype.Groq:
+		return "groq", groq.ModelList
+	case channeltype.LingYiWanWu:
+		return "lingyiwanwu", lingyiwanwu.ModelList
+	case channeltype.StepFun:
+		return "stepfun", stepfun.ModelList
+	case channeltype.DeepSeek:
+		return "deepseek", deepseek.ModelList
+	case channeltype.TogetherAI:
+		return "together.ai", togetherai.ModelList
+	case channeltype.Doubao:
+		return "doubao", doubao.ModelList
+	case channeltype.Novita:
+		return "novita", novita.ModelList
+	case channeltype.SiliconFlow:
+		return "siliconflow", siliconflow.ModelList
+	case channeltype.XAI:
+		return "xai", xai.ModelList
+	case channeltype.BaiduV2:
+		return "baiduv2", baiduv2.ModelList
+	case channeltype.XunfeiV2:
+		return "xunfeiv2", xunfeiv2.ModelList
+	case channeltype.OpenRouter:
+		return "openrouter", openrouter.ModelList
+	case channeltype.AliBailian:
+		return "alibailian", alibailian.ModelList
+	case channeltype.GeminiOpenAICompatible:
+		return "geminiv2", geminiv2.ModelList
+	default:
+		return "openai", ModelList
+	}
+}
--- a/relay/adaptor/openai/constants.go
+++ b/relay/adaptor/openai/constants.go
@@ -0,0 +1,27 @@
+package openai
+
+var ModelList = []string{
+	"gpt-3.5-turbo", "gpt-3.5-turbo-0301", "gpt-3.5-turbo-0613", "gpt-3.5-turbo-1106", "gpt-3.5-turbo-0125",
+	"gpt-3.5-turbo-16k", "gpt-3.5-turbo-16k-0613",
+	"gpt-3.5-turbo-instruct",
+	"gpt-4", "gpt-4-0314", "gpt-4-0613", "gpt-4-1106-preview", "gpt-4-0125-preview",
+	"gpt-4-32k", "gpt-4-32k-0314", "gpt-4-32k-0613",
+	"gpt-4-turbo-preview", "gpt-4-turbo", "gpt-4-turbo-2024-04-09",
+	"gpt-4o", "gpt-4o-2024-05-13",
+	"gpt-4o-2024-08-06",
+	"gpt-4o-2024-11-20",
+	"chatgpt-4o-latest",
+	"gpt-4o-mini", "gpt-4o-mini-2024-07-18",
+	"gpt-4-vision-preview",
+	"text-embedding-ada-002", "text-embedding-3-small", "text-embedding-3-large",
+	"text-curie-001", "text-babbage-001", "text-ada-001", "text-davinci-002", "text-davinci-003",
+	"text-moderation-latest", "text-moderation-stable",
+	"text-davinci-edit-001",
+	"davinci-002", "babbage-002",
+	"dall-e-2", "dall-e-3",
+	"whisper-1",
+	"tts-1", "tts-1-1106", "tts-1-hd", "tts-1-hd-1106",
+	"o1", "o1-2024-12-17",
+	"o1-preview", "o1-preview-2024-09-12",
+	"o1-mini", "o1-mini-2024-09-12",
+}
--- a/relay/adaptor/openai/helper.go
+++ b/relay/adaptor/openai/helper.go
@@ -0,0 +1,34 @@
+package openai
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/songquanpeng/one-api/relay/channeltype"
+	"github.com/songquanpeng/one-api/relay/model"
+)
+
+func ResponseText2Usage(responseText string, modelName string, promptTokens int) *model.Usage {
+	usage := &model.Usage{}
+	usage.PromptTokens = promptTokens
+	usage.CompletionTokens = CountTokenText(responseText, modelName)
+	usage.TotalTokens = usage.PromptTokens + usage.CompletionTokens
+	return usage
+}
+
+func GetFullRequestURL(baseURL string, requestURL string, channelType int) string {
+	if channelType == channeltype.OpenAICompatible {
+		return fmt.Sprintf("%s%s", strings.TrimSuffix(baseURL, "/"), strings.TrimPrefix(requestURL, "/v1"))
+	}
+	fullRequestURL := fmt.Sprintf("%s%s", baseURL, requestURL)
+
+	if strings.HasPrefix(baseURL, "https://gateway.ai.cloudflare.com") {
+		switch channelType {
+		case channeltype.OpenAI:
+			fullRequestURL = fmt.Sprintf("%s%s", baseURL, strings.TrimPrefix(requestURL, "/v1"))
+		case channeltype.Azure:
+			fullRequestURL = fmt.Sprintf("%s%s", baseURL, strings.TrimPrefix(requestURL, "/openai/deployments"))
+		}
+	}
+	return fullRequestURL
+}
--- a/relay/adaptor/openai/image.go
+++ b/relay/adaptor/openai/image.go
@@ -0,0 +1,44 @@
+package openai
+
+import (
+	"bytes"
+	"encoding/json"
+	"github.com/gin-gonic/gin"
+	"github.com/songquanpeng/one-api/relay/model"
+	"io"
+	"net/http"
+)
+
+func ImageHandler(c *gin.Context, resp *http.Response) (*model.ErrorWithStatusCode, *model.Usage) {
+	var imageResponse ImageResponse
+	responseBody, err := io.ReadAll(resp.Body)
+
+	if err != nil {
+		return ErrorWrapper(err, "read_response_body_failed", http.StatusInternalServerError), nil
+	}
+	err = resp.Body.Close()
+	if err != nil {
+		return ErrorWrapper(err, "close_response_body_failed", http.StatusInternalServerError), nil
+	}
+	err = json.Unmarshal(responseBody, &imageResponse)
+	if err != nil {
+		return ErrorWrapper(err, "unmarshal_response_body_failed", http.StatusInternalServerError), nil
+	}
+
+	resp.Body = io.NopCloser(bytes.NewBuffer(responseBody))
+
+	for k, v := range resp.Header {
+		c.Writer.Header().Set(k, v[0])
+	}
+	c.Writer.WriteHeader(resp.StatusCode)
+
+	_, err = io.Copy(c.Writer, resp.Body)
+	if err != nil {
+		return ErrorWrapper(err, "copy_response_body_failed", http.StatusInternalServerError), nil
+	}
+	err = resp.Body.Close()
+	if err != nil {
+		return ErrorWrapper(err, "close_response_body_failed", http.StatusInternalServerError), nil
+	}
+	return nil, nil
+}
--- a/relay/adaptor/openai/main.go
+++ b/relay/adaptor/openai/main.go
@@ -0,0 +1,151 @@
+package openai
+
+import (
+	"bufio"
+	"bytes"
+	"encoding/json"
+	"io"
+	"net/http"
+	"strings"
+
+	"github.com/songquanpeng/one-api/common/render"
+
+	"github.com/gin-gonic/gin"
+	"github.com/songquanpeng/one-api/common"
+	"github.com/songquanpeng/one-api/common/conv"
+	"github.com/songquanpeng/one-api/common/logger"
+	"github.com/songquanpeng/one-api/relay/model"
+	"github.com/songquanpeng/one-api/relay/relaymode"
+)
+
+const (
+	dataPrefix       = "data: "
+	done             = "[DONE]"
+	dataPrefixLength = len(dataPrefix)
+)
+
+func StreamHandler(c *gin.Context, resp *http.Response, relayMode int) (*model.ErrorWithStatusCode, string, *model.Usage) {
+	responseText := ""
+	scanner := bufio.NewScanner(resp.Body)
+	scanner.Split(bufio.ScanLines)
+	var usage *model.Usage
+
+	common.SetEventStreamHeaders(c)
+
+	doneRendered := false
+	for scanner.Scan() {
+		data := scanner.Text()
+		if len(data) < dataPrefixLength { // ignore blank line or wrong format
+			continue
+		}
+		if data[:dataPrefixLength] != dataPrefix && data[:dataPrefixLength] != done {
+			continue
+		}
+		if strings.HasPrefix(data[dataPrefixLength:], done) {
+			render.StringData(c, data)
+			doneRendered = true
+			continue
+		}
+		switch relayMode {
+		case relaymode.ChatCompletions:
+			var streamResponse ChatCompletionsStreamResponse
+			err := json.Unmarshal([]byte(data[dataPrefixLength:]), &streamResponse)
+			if err != nil {
+				logger.SysError("error unmarshalling stream response: " + err.Error())
+				render.StringData(c, data) // if error happened, pass the data to client
+				continue                   // just ignore the error
+			}
+			if len(streamResponse.Choices) == 0 && streamResponse.Usage == nil {
+				// but for empty choice and no usage, we should not pass it to client, this is for azure
+				continue // just ignore empty choice
+			}
+			render.StringData(c, data)
+			for _, choice := range streamResponse.Choices {
+				responseText += conv.AsString(choice.Delta.Content)
+			}
+			if streamResponse.Usage != nil {
+				usage = streamResponse.Usage
+			}
+		case relaymode.Completions:
+			render.StringData(c, data)
+			var streamResponse CompletionsStreamResponse
+			err := json.Unmarshal([]byte(data[dataPrefixLength:]), &streamResponse)
+			if err != nil {
+				logger.SysError("error unmarshalling stream response: " + err.Error())
+				continue
+			}
+			for _, choice := range streamResponse.Choices {
+				responseText += choice.Text
+			}
+		}
+	}
+
+	if err := scanner.Err(); err != nil {
+		logger.SysError("error reading stream: " + err.Error())
+	}
+
+	if !doneRendered {
+		render.Done(c)
+	}
+
+	err := resp.Body.Close()
+	if err != nil {
+		return ErrorWrapper(err, "close_response_body_failed", http.StatusInternalServerError), "", nil
+	}
+
+	return nil, responseText, usage
+}
+
+func Handler(c *gin.Context, resp *http.Response, promptTokens int, modelName string) (*model.ErrorWithStatusCode, *model.Usage) {
+	var textResponse SlimTextResponse
+	responseBody, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return ErrorWrapper(err, "read_response_body_failed", http.StatusInternalServerError), nil
+	}
+	err = resp.Body.Close()
+	if err != nil {
+		return ErrorWrapper(err, "close_response_body_failed", http.StatusInternalServerError), nil
+	}
+	err = json.Unmarshal(responseBody, &textResponse)
+	if err != nil {
+		return ErrorWrapper(err, "unmarshal_response_body_failed", http.StatusInternalServerError), nil
+	}
+	if textResponse.Error.Type != "" {
+		return &model.ErrorWithStatusCode{
+			Error:      textResponse.Error,
+			StatusCode: resp.StatusCode,
+		}, nil
+	}
+	// Reset response body
+	resp.Body = io.NopCloser(bytes.NewBuffer(responseBody))
+
+	// We shouldn't set the header before we parse the response body, because the parse part may fail.
+	// And then we will have to send an error response, but in this case, the header has already been set.
+	// So the HTTPClient will be confused by the response.
+	// For example, Postman will report error, and we cannot check the response at all.
+	for k, v := range resp.Header {
+		c.Writer.Header().Set(k, v[0])
+	}
+	c.Writer.WriteHeader(resp.StatusCode)
+	_, err = io.Copy(c.Writer, resp.Body)
+	if err != nil {
+		return ErrorWrapper(err, "copy_response_body_failed", http.StatusInternalServerError), nil
+	}
+	err = resp.Body.Close()
+	if err != nil {
+		return ErrorWrapper(err, "close_response_body_failed", http.StatusInternalServerError), nil
+	}
+
+	if textResponse.Usage.TotalTokens == 0 || (textResponse.Usage.PromptTokens == 0 && textResponse.Usage.CompletionTokens == 0) {
+		completionTokens := 0
+		for _, choice := range textResponse.Choices {
+			completionTokens += CountTokenText(choice.Message.StringContent(), modelName)
+		}
+		textResponse.Usage = model.Usage{
+			PromptTokens:     promptTokens,
+			CompletionTokens: completionTokens,
+			TotalTokens:      promptTokens + completionTokens,
+		}
+	}
+	return nil, &textResponse.Usage
+}
--- a/relay/adaptor/openai/model.go
+++ b/relay/adaptor/openai/model.go
@@ -0,0 +1,145 @@
+package openai
+
+import "github.com/songquanpeng/one-api/relay/model"
+
+type TextContent struct {
+	Type string `json:"type,omitempty"`
+	Text string `json:"text,omitempty"`
+}
+
+type ImageContent struct {
+	Type     string          `json:"type,omitempty"`
+	ImageURL *model.ImageURL `json:"image_url,omitempty"`
+}
+
+type ChatRequest struct {
+	Model     string          `json:"model"`
+	Messages  []model.Message `json:"messages"`
+	MaxTokens int             `json:"max_tokens"`
+}
+
+type TextRequest struct {
+	Model     string          `json:"model"`
+	Messages  []model.Message `json:"messages"`
+	Prompt    string          `json:"prompt"`
+	MaxTokens int             `json:"max_tokens"`
+	//Stream   bool      `json:"stream"`
+}
+
+// ImageRequest docs: https://platform.openai.com/docs/api-reference/images/create
+type ImageRequest struct {
+	Model          string `json:"model"`
+	Prompt         string `json:"prompt" binding:"required"`
+	N              int    `json:"n,omitempty"`
+	Size           string `json:"size,omitempty"`
+	Quality        string `json:"quality,omitempty"`
+	ResponseFormat string `json:"response_format,omitempty"`
+	Style          string `json:"style,omitempty"`
+	User           string `json:"user,omitempty"`
+}
+
+type WhisperJSONResponse struct {
+	Text string `json:"text,omitempty"`
+}
+
+type WhisperVerboseJSONResponse struct {
+	Task     string    `json:"task,omitempty"`
+	Language string    `json:"language,omitempty"`
+	Duration float64   `json:"duration,omitempty"`
+	Text     string    `json:"text,omitempty"`
+	Segments []Segment `json:"segments,omitempty"`
+}
+
+type Segment struct {
+	Id               int     `json:"id"`
+	Seek             int     `json:"seek"`
+	Start            float64 `json:"start"`
+	End              float64 `json:"end"`
+	Text             string  `json:"text"`
+	Tokens           []int   `json:"tokens"`
+	Temperature      float64 `json:"temperature"`
+	AvgLogprob       float64 `json:"avg_logprob"`
+	CompressionRatio float64 `json:"compression_ratio"`
+	NoSpeechProb     float64 `json:"no_speech_prob"`
+}
+
+type TextToSpeechRequest struct {
+	Model          string  `json:"model" binding:"required"`
+	Input          string  `json:"input" binding:"required"`
+	Voice          string  `json:"voice" binding:"required"`
+	Speed          float64 `json:"speed"`
+	ResponseFormat string  `json:"response_format"`
+}
+
+type UsageOrResponseText struct {
+	*model.Usage
+	ResponseText string
+}
+
+type SlimTextResponse struct {
+	Choices     []TextResponseChoice `json:"choices"`
+	model.Usage `json:"usage"`
+	Error       model.Error `json:"error"`
+}
+
+type TextResponseChoice struct {
+	Index         int `json:"index"`
+	model.Message `json:"message"`
+	FinishReason  string `json:"finish_reason"`
+}
+
+type TextResponse struct {
+	Id          string               `json:"id"`
+	Model       string               `json:"model,omitempty"`
+	Object      string               `json:"object"`
+	Created     int64                `json:"created"`
+	Choices     []TextResponseChoice `json:"choices"`
+	model.Usage `json:"usage"`
+}
+
+type EmbeddingResponseItem struct {
+	Object    string    `json:"object"`
+	Index     int       `json:"index"`
+	Embedding []float64 `json:"embedding"`
+}
+
+type EmbeddingResponse struct {
+	Object      string                  `json:"object"`
+	Data        []EmbeddingResponseItem `json:"data"`
+	Model       string                  `json:"model"`
+	model.Usage `json:"usage"`
+}
+
+type ImageData struct {
+	Url           string `json:"url,omitempty"`
+	B64Json       string `json:"b64_json,omitempty"`
+	RevisedPrompt string `json:"revised_prompt,omitempty"`
+}
+
+type ImageResponse struct {
+	Created int64       `json:"created"`
+	Data    []ImageData `json:"data"`
+	//model.Usage `json:"usage"`
+}
+
+type ChatCompletionsStreamResponseChoice struct {
+	Index        int           `json:"index"`
+	Delta        model.Message `json:"delta"`
+	FinishReason *string       `json:"finish_reason,omitempty"`
+}
+
+type ChatCompletionsStreamResponse struct {
+	Id      string                                `json:"id"`
+	Object  string                                `json:"object"`
+	Created int64                                 `json:"created"`
+	Model   string                                `json:"model"`
+	Choices []ChatCompletionsStreamResponseChoice `json:"choices"`
+	Usage   *model.Usage                          `json:"usage,omitempty"`
+}
+
+type CompletionsStreamResponse struct {
+	Choices []struct {
+		Text         string `json:"text"`
+		FinishReason string `json:"finish_reason"`
+	} `json:"choices"`
+}
--- a/relay/adaptor/openai/token.go
+++ b/relay/adaptor/openai/token.go
@@ -0,0 +1,235 @@
+package openai
+
+import (
+	"errors"
+	"fmt"
+	"math"
+	"strings"
+
+	"github.com/pkoukk/tiktoken-go"
+
+	"github.com/songquanpeng/one-api/common/config"
+	"github.com/songquanpeng/one-api/common/image"
+	"github.com/songquanpeng/one-api/common/logger"
+	billingratio "github.com/songquanpeng/one-api/relay/billing/ratio"
+	"github.com/songquanpeng/one-api/relay/model"
+)
+
+// tokenEncoderMap won't grow after initialization
+var tokenEncoderMap = map[string]*tiktoken.Tiktoken{}
+var defaultTokenEncoder *tiktoken.Tiktoken
+
+func InitTokenEncoders() {
+	logger.SysLog("initializing token encoders")
+	gpt35TokenEncoder, err := tiktoken.EncodingForModel("gpt-3.5-turbo")
+	if err != nil {
+		logger.FatalLog(fmt.Sprintf("failed to get gpt-3.5-turbo token encoder: %s, "+
+			"if you are using in offline environment, please set TIKTOKEN_CACHE_DIR to use exsited files, check this link for more information: https://stackoverflow.com/questions/76106366/how-to-use-tiktoken-in-offline-mode-computer ", err.Error()))
+	}
+	defaultTokenEncoder = gpt35TokenEncoder
+	gpt4oTokenEncoder, err := tiktoken.EncodingForModel("gpt-4o")
+	if err != nil {
+		logger.FatalLog(fmt.Sprintf("failed to get gpt-4o token encoder: %s", err.Error()))
+	}
+	gpt4TokenEncoder, err := tiktoken.EncodingForModel("gpt-4")
+	if err != nil {
+		logger.FatalLog(fmt.Sprintf("failed to get gpt-4 token encoder: %s", err.Error()))
+	}
+	for model := range billingratio.ModelRatio {
+		if strings.HasPrefix(model, "gpt-3.5") {
+			tokenEncoderMap[model] = gpt35TokenEncoder
+		} else if strings.HasPrefix(model, "gpt-4o") {
+			tokenEncoderMap[model] = gpt4oTokenEncoder
+		} else if strings.HasPrefix(model, "gpt-4") {
+			tokenEncoderMap[model] = gpt4TokenEncoder
+		} else {
+			tokenEncoderMap[model] = nil
+		}
+	}
+	logger.SysLog("token encoders initialized")
+}
+
+func getTokenEncoder(model string) *tiktoken.Tiktoken {
+	tokenEncoder, ok := tokenEncoderMap[model]
+	if ok && tokenEncoder != nil {
+		return tokenEncoder
+	}
+	if ok {
+		tokenEncoder, err := tiktoken.EncodingForModel(model)
+		if err != nil {
+			logger.SysError(fmt.Sprintf("failed to get token encoder for model %s: %s, using encoder for gpt-3.5-turbo", model, err.Error()))
+			tokenEncoder = defaultTokenEncoder
+		}
+		tokenEncoderMap[model] = tokenEncoder
+		return tokenEncoder
+	}
+	return defaultTokenEncoder
+}
+
+func getTokenNum(tokenEncoder *tiktoken.Tiktoken, text string) int {
+	if config.ApproximateTokenEnabled {
+		return int(float64(len(text)) * 0.38)
+	}
+	return len(tokenEncoder.Encode(text, nil, nil))
+}
+
+func CountTokenMessages(messages []model.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 model == "gpt-3.5-turbo-0301" {
+		tokensPerMessage = 4
+		tokensPerName = -1 // If there's a name, the role is omitted
+	} else {
+		tokensPerMessage = 3
+		tokensPerName = 1
+	}
+	tokenNum := 0
+	for _, message := range messages {
+		tokenNum += tokensPerMessage
+		switch v := message.Content.(type) {
+		case string:
+			tokenNum += getTokenNum(tokenEncoder, v)
+		case []any:
+			for _, it := range v {
+				m := it.(map[string]any)
+				switch m["type"] {
+				case "text":
+					if textValue, ok := m["text"]; ok {
+						if textString, ok := textValue.(string); ok {
+							tokenNum += getTokenNum(tokenEncoder, textString)
+						}
+					}
+				case "image_url":
+					imageUrl, ok := m["image_url"].(map[string]any)
+					if ok {
+						url := imageUrl["url"].(string)
+						detail := ""
+						if imageUrl["detail"] != nil {
+							detail = imageUrl["detail"].(string)
+						}
+						imageTokens, err := countImageTokens(url, detail, model)
+						if err != nil {
+							logger.SysError("error counting image tokens: " + err.Error())
+						} else {
+							tokenNum += imageTokens
+						}
+					}
+				}
+			}
+		}
+		tokenNum += getTokenNum(tokenEncoder, message.Role)
+		if message.Name != nil {
+			tokenNum += tokensPerName
+			tokenNum += getTokenNum(tokenEncoder, *message.Name)
+		}
+	}
+	tokenNum += 3 // Every reply is primed with <|start|>assistant<|message|>
+	return tokenNum
+}
+
+const (
+	lowDetailCost         = 85
+	highDetailCostPerTile = 170
+	additionalCost        = 85
+	// gpt-4o-mini cost higher than other model
+	gpt4oMiniLowDetailCost  = 2833
+	gpt4oMiniHighDetailCost = 5667
+	gpt4oMiniAdditionalCost = 2833
+)
+
+// https://platform.openai.com/docs/guides/vision/calculating-costs
+// https://github.com/openai/openai-cookbook/blob/05e3f9be4c7a2ae7ecf029a7c32065b024730ebe/examples/How_to_count_tokens_with_tiktoken.ipynb
+func countImageTokens(url string, detail string, model string) (_ int, err error) {
+	var fetchSize = true
+	var width, height int
+	// Reference: https://platform.openai.com/docs/guides/vision/low-or-high-fidelity-image-understanding
+	// detail == "auto" is undocumented on how it works, it just said the model will use the auto setting which will look at the image input size and decide if it should use the low or high setting.
+	// According to the official guide, "low" disable the high-res model,
+	// and only receive low-res 512px x 512px version of the image, indicating
+	// that image is treated as low-res when size is smaller than 512px x 512px,
+	// then we can assume that image size larger than 512px x 512px is treated
+	// as high-res. Then we have the following logic:
+	// if detail == "" || detail == "auto" {
+	// 	width, height, err = image.GetImageSize(url)
+	// 	if err != nil {
+	// 		return 0, err
+	// 	}
+	// 	fetchSize = false
+	// 	// not sure if this is correct
+	// 	if width > 512 || height > 512 {
+	// 		detail = "high"
+	// 	} else {
+	// 		detail = "low"
+	// 	}
+	// }
+
+	// However, in my test, it seems to be always the same as "high".
+	// The following image, which is 125x50, is still treated as high-res, taken
+	// 255 tokens in the response of non-stream chat completion api.
+	// https://upload.wikimedia.org/wikipedia/commons/1/10/18_Infantry_Division_Messina.jpg
+	if detail == "" || detail == "auto" {
+		// assume by test, not sure if this is correct
+		detail = "high"
+	}
+	switch detail {
+	case "low":
+		if strings.HasPrefix(model, "gpt-4o-mini") {
+			return gpt4oMiniLowDetailCost, nil
+		}
+		return lowDetailCost, nil
+	case "high":
+		if fetchSize {
+			width, height, err = image.GetImageSize(url)
+			if err != nil {
+				return 0, err
+			}
+		}
+		if width > 2048 || height > 2048 { // max(width, height) > 2048
+			ratio := float64(2048) / math.Max(float64(width), float64(height))
+			width = int(float64(width) * ratio)
+			height = int(float64(height) * ratio)
+		}
+		if width > 768 && height > 768 { // min(width, height) > 768
+			ratio := float64(768) / math.Min(float64(width), float64(height))
+			width = int(float64(width) * ratio)
+			height = int(float64(height) * ratio)
+		}
+		numSquares := int(math.Ceil(float64(width)/512) * math.Ceil(float64(height)/512))
+		if strings.HasPrefix(model, "gpt-4o-mini") {
+			return numSquares*gpt4oMiniHighDetailCost + gpt4oMiniAdditionalCost, nil
+		}
+		result := numSquares*highDetailCostPerTile + additionalCost
+		return result, nil
+	default:
+		return 0, errors.New("invalid detail option")
+	}
+}
+
+func CountTokenInput(input any, model string) int {
+	switch v := input.(type) {
+	case string:
+		return CountTokenText(v, model)
+	case []string:
+		text := ""
+		for _, s := range v {
+			text += s
+		}
+		return CountTokenText(text, model)
+	}
+	return 0
+}
+
+func CountTokenText(text string, model string) int {
+	tokenEncoder := getTokenEncoder(model)
+	return getTokenNum(tokenEncoder, text)
+}
+
+func CountToken(text string) int {
+	return CountTokenInput(text, "gpt-3.5-turbo")
+}
--- a/relay/adaptor/openai/util.go
+++ b/relay/adaptor/openai/util.go
@@ -0,0 +1,23 @@
+package openai
+
+import (
+	"context"
+	"fmt"
+
+	"github.com/songquanpeng/one-api/common/logger"
+	"github.com/songquanpeng/one-api/relay/model"
+)
+
+func ErrorWrapper(err error, code string, statusCode int) *model.ErrorWithStatusCode {
+	logger.Error(context.TODO(), fmt.Sprintf("[%s]%+v", code, err))
+
+	Error := model.Error{
+		Message: err.Error(),
+		Type:    "one_api_error",
+		Code:    code,
+	}
+	return &model.ErrorWithStatusCode{
+		Error:      Error,
+		StatusCode: statusCode,
+	}
+}