✨ feat: Support Claude3 (#86)

* Claude改用messages API,支持Claude3 * 删除新API不支持的模型 * 忘了改请求地址 * 🐛 fix: fix the problem of return format and completion token not being obtained --------- Co-authored-by: MartialBE <me@xiao5.info>
2024-03-07 01:21:07 +08:00 · 2024-03-07 01:21:07 +08:00 · e1fcfae928
commit e1fcfae928
parent 07c18dfb91
6 changed files with 131 additions and 75 deletions
--- a/common/model-ratio.go
+++ b/common/model-ratio.go
@ -87,11 +87,12 @@ func init() {
 		"dall-e-3": {[]float64{20, 20}, ChannelTypeOpenAI},

 		// $0.80/million tokens $2.40/million tokens
-		"claude-instant-1": {[]float64{0.4, 1.2}, ChannelTypeAnthropic},
+		"claude-instant-1.2": {[]float64{0.4, 1.2}, ChannelTypeAnthropic},
 		// $8.00/million tokens $24.00/million tokens
-		"claude-2":   {[]float64{4, 12}, ChannelTypeAnthropic},
-		"claude-2.0": {[]float64{4, 12}, ChannelTypeAnthropic},
-		"claude-2.1": {[]float64{4, 12}, ChannelTypeAnthropic},
+		"claude-2.0":               {[]float64{4, 12}, ChannelTypeAnthropic},
+		"claude-2.1":               {[]float64{4, 12}, ChannelTypeAnthropic},
+		"claude-3-opus-20240229":   {[]float64{7.5, 22.5}, ChannelTypeAnthropic},
+		"claude-3-sonnet-20240229": {[]float64{1.3, 3.9}, ChannelTypeAnthropic},

 		// ￥0.012 / 1k tokens ￥0.012 / 1k tokens
 		"ERNIE-Bot": {[]float64{0.8572, 0.8572}, ChannelTypeBaidu},
@ -291,7 +292,7 @@ func GetCompletionRatio(name string) float64 {
 		}
 		return 2
 	}
-	if strings.HasPrefix(name, "claude-instant-1") {
+	if strings.HasPrefix(name, "claude-instant-1.2") {
 		return 3.38
 	}
 	if strings.HasPrefix(name, "claude-2") {
--- a/modelRatio.json
+++ b/modelRatio.json
@ -43,10 +43,11 @@
  "text-moderation-latest": [0.1, 0.1],
  "dall-e-2": [8, 8],
  "dall-e-3": [20, 20],
-  "claude-instant-1": [0.4, 1.2],
-  "claude-2": [4, 12],
+  "claude-instant-1.2": [0.4, 1.2],
  "claude-2.0": [4, 12],
  "claude-2.1": [4, 12],
+  "claude-3-opus-20240229": [7.5, 22.5],
+  "claude-3-sonnet-20240229": [1.3, 3.9],
  "ERNIE-Bot": [0.8572, 0.8572],
  "ERNIE-Bot-8k": [1.7143, 3.4286],
  "ERNIE-Bot-turbo": [0.5715, 0.5715],
--- a/providers/claude/base.go
+++ b/providers/claude/base.go
@ -29,13 +29,13 @@ type ClaudeProvider struct {
 func getConfig() base.ProviderConfig {
 	return base.ProviderConfig{
 		BaseURL:         "https://api.anthropic.com",
-		ChatCompletions: "/v1/complete",
+		ChatCompletions: "/v1/messages",
 	}
 }

 // 请求错误处理
 func requestErrorHandle(resp *http.Response) *types.OpenAIError {
-	claudeError := &ClaudeResponseError{}
+	claudeError := &ClaudeError{}
 	err := json.NewDecoder(resp.Body).Decode(claudeError)
 	if err != nil {
 		return nil
@ -45,14 +45,14 @@ func requestErrorHandle(resp *http.Response) *types.OpenAIError {
 }

 // 错误处理
-func errorHandle(claudeError *ClaudeResponseError) *types.OpenAIError {
-	if claudeError.Error.Type == "" {
+func errorHandle(claudeError *ClaudeError) *types.OpenAIError {
+	if claudeError.Type == "" {
 		return nil
 	}
 	return &types.OpenAIError{
-		Message: claudeError.Error.Message,
-		Type:    claudeError.Error.Type,
-		Code:    claudeError.Error.Type,
+		Message: claudeError.Message,
+		Type:    claudeError.Type,
+		Code:    claudeError.Type,
 	}
 }

@ -73,7 +73,7 @@ func (p *ClaudeProvider) GetRequestHeaders() (headers map[string]string) {

 func stopReasonClaude2OpenAI(reason string) string {
 	switch reason {
-	case "stop_sequence":
+	case "end_turn":
 		return types.FinishReasonStop
 	case "max_tokens":
 		return types.FinishReasonLength
--- a/providers/claude/chat.go
+++ b/providers/claude/chat.go
@ -83,36 +83,36 @@ func (p *ClaudeProvider) getChatRequest(request *types.ChatCompletionRequest) (*

 func convertFromChatOpenai(request *types.ChatCompletionRequest) *ClaudeRequest {
 	claudeRequest := ClaudeRequest{
-		Model:             request.Model,
-		Prompt:            "",
-		MaxTokensToSample: request.MaxTokens,
-		StopSequences:     nil,
-		Temperature:       request.Temperature,
-		TopP:              request.TopP,
-		Stream:            request.Stream,
+		Model:         request.Model,
+		Messages:      nil,
+		System:        "",
+		MaxTokens:     request.MaxTokens,
+		StopSequences: nil,
+		Temperature:   request.Temperature,
+		TopP:          request.TopP,
+		Stream:        request.Stream,
 	}
-	if claudeRequest.MaxTokensToSample == 0 {
-		claudeRequest.MaxTokensToSample = 1000000
+	if claudeRequest.MaxTokens == 0 {
+		claudeRequest.MaxTokens = 4096
 	}
-	prompt := ""
+	var messages []Message
 	for _, message := range request.Messages {
-		if message.Role == "user" {
-			prompt += fmt.Sprintf("\n\nHuman: %s", message.Content)
-		} else if message.Role == "assistant" {
-			prompt += fmt.Sprintf("\n\nAssistant: %s", message.Content)
-		} else if message.Role == "system" {
-			if prompt == "" {
-				prompt = message.StringContent()
-			}
+		if message.Role != "system" {
+			messages = append(messages, Message{
+				Role:    message.Role,
+				Content: message.Content.(string),
+			})
+			claudeRequest.Messages = messages
+		} else {
+			claudeRequest.System = message.Content.(string)
 		}
 	}
-	prompt += "\n\nAssistant:"
-	claudeRequest.Prompt = prompt
+
 	return &claudeRequest
 }

 func (p *ClaudeProvider) convertToChatOpenai(response *ClaudeResponse, request *types.ChatCompletionRequest) (openaiResponse *types.ChatCompletionResponse, errWithCode *types.OpenAIErrorWithStatusCode) {
-	error := errorHandle(&response.ClaudeResponseError)
+	error := errorHandle(&response.Error)
 	if error != nil {
 		errWithCode = &types.OpenAIErrorWithStatusCode{
 			OpenAIError: *error,
@ -125,26 +125,33 @@ func (p *ClaudeProvider) convertToChatOpenai(response *ClaudeResponse, request *
 		Index: 0,
 		Message: types.ChatCompletionMessage{
 			Role:    "assistant",
-			Content: strings.TrimPrefix(response.Completion, " "),
+			Content: strings.TrimPrefix(response.Content[0].Text, " "),
 			Name:    nil,
 		},
 		FinishReason: stopReasonClaude2OpenAI(response.StopReason),
 	}
 	openaiResponse = &types.ChatCompletionResponse{
-		ID:      fmt.Sprintf("chatcmpl-%s", common.GetUUID()),
+		ID:      response.Id,
 		Object:  "chat.completion",
 		Created: common.GetTimestamp(),
 		Choices: []types.ChatCompletionChoice{choice},
 		Model:   response.Model,
+		Usage: &types.Usage{
+			CompletionTokens: 0,
+			PromptTokens:     0,
+			TotalTokens:      0,
+		},
 	}

-	completionTokens := common.CountTokenText(response.Completion, response.Model)
-	response.Usage.CompletionTokens = completionTokens
-	response.Usage.TotalTokens = response.Usage.PromptTokens + completionTokens
+	completionTokens := response.Usage.OutputTokens

-	openaiResponse.Usage = response.Usage
+	promptTokens := response.Usage.InputTokens

-	*p.Usage = *response.Usage
+	openaiResponse.Usage.PromptTokens = promptTokens
+	openaiResponse.Usage.CompletionTokens = completionTokens
+	openaiResponse.Usage.TotalTokens = promptTokens + completionTokens
+
+	*p.Usage = *openaiResponse.Usage

 	return openaiResponse, nil
 }
@ -152,7 +159,7 @@ func (p *ClaudeProvider) convertToChatOpenai(response *ClaudeResponse, request *
 // 转换为OpenAI聊天流式请求体
 func (h *claudeStreamHandler) handlerStream(rawLine *[]byte, dataChan chan string, errChan chan error) {
 	// 如果rawLine 前缀不为data:，则直接返回
-	if !strings.HasPrefix(string(*rawLine), `data: {"type": "completion"`) {
+	if !strings.HasPrefix(string(*rawLine), `data: {"type"`) {
 		*rawLine = nil
 		return
 	}
@ -160,43 +167,61 @@ func (h *claudeStreamHandler) handlerStream(rawLine *[]byte, dataChan chan strin
 	// 去除前缀
 	*rawLine = (*rawLine)[6:]

-	var claudeResponse *ClaudeResponse
-	err := json.Unmarshal(*rawLine, claudeResponse)
+	var claudeResponse ClaudeStreamResponse
+	err := json.Unmarshal(*rawLine, &claudeResponse)
 	if err != nil {
 		errChan <- common.ErrorToOpenAIError(err)
 		return
 	}

-	error := errorHandle(&claudeResponse.ClaudeResponseError)
+	error := errorHandle(&claudeResponse.Error)
 	if error != nil {
 		errChan <- error
 		return
 	}

-	if claudeResponse.StopReason == "stop_sequence" {
+	switch claudeResponse.Type {
+	case "message_start":
+		h.Usage.PromptTokens = claudeResponse.Message.InputTokens
+
+	case "message_delta":
+		h.convertToOpenaiStream(&claudeResponse, dataChan, errChan)
+		h.Usage.CompletionTokens = claudeResponse.Usage.OutputTokens
+		h.Usage.TotalTokens = h.Usage.PromptTokens + h.Usage.CompletionTokens
+
+	case "content_block_delta":
+		h.convertToOpenaiStream(&claudeResponse, dataChan, errChan)
+
+	case "message_stop":
 		errChan <- io.EOF
 		*rawLine = requester.StreamClosed
+
+	default:
 		return
 	}
-
-	h.convertToOpenaiStream(claudeResponse, dataChan, errChan)
 }

-func (h *claudeStreamHandler) convertToOpenaiStream(claudeResponse *ClaudeResponse, dataChan chan string, errChan chan error) {
-	var choice types.ChatCompletionStreamChoice
-	choice.Delta.Content = claudeResponse.Completion
-	finishReason := stopReasonClaude2OpenAI(claudeResponse.StopReason)
-	if finishReason != "null" {
+func (h *claudeStreamHandler) convertToOpenaiStream(claudeResponse *ClaudeStreamResponse, dataChan chan string, errChan chan error) {
+	choice := types.ChatCompletionStreamChoice{
+		Index: claudeResponse.Index,
+	}
+
+	if claudeResponse.Delta.Text != "" {
+		choice.Delta.Content = claudeResponse.Delta.Text
+	}
+
+	finishReason := stopReasonClaude2OpenAI(claudeResponse.Delta.StopReason)
+	if finishReason != "" {
 		choice.FinishReason = &finishReason
 	}
 	chatCompletion := types.ChatCompletionStreamResponse{
+		ID:      fmt.Sprintf("chatcmpl-%s", common.GetUUID()),
 		Object:  "chat.completion.chunk",
+		Created: common.GetTimestamp(),
 		Model:   h.Request.Model,
 		Choices: []types.ChatCompletionStreamChoice{choice},
 	}

 	responseBody, _ := json.Marshal(chatCompletion)
 	dataChan <- string(responseBody)
-
-	h.Usage.PromptTokens += common.CountTokenText(claudeResponse.Completion, h.Request.Model)
 }
--- a/providers/claude/type.go
+++ b/providers/claude/type.go
@ -1,7 +1,5 @@
 package claude

-import "one-api/types"
-
 type ClaudeError struct {
 	Type    string `json:"type"`
 	Message string `json:"message"`
@ -11,25 +9,56 @@ type ClaudeMetadata struct {
 	UserId string `json:"user_id"`
 }

+type ResContent struct {
+	Text string `json:"text"`
+	Type string `json:"type"`
+}
+
+type Message struct {
+	Role    string `json:"role"`
+	Content string `json:"content"`
+}
+
 type ClaudeRequest struct {
-	Model             string   `json:"model"`
-	Prompt            string   `json:"prompt"`
-	MaxTokensToSample int      `json:"max_tokens_to_sample"`
-	StopSequences     []string `json:"stop_sequences,omitempty"`
-	Temperature       float64  `json:"temperature,omitempty"`
-	TopP              float64  `json:"top_p,omitempty"`
-	TopK              int      `json:"top_k,omitempty"`
+	Model         string    `json:"model"`
+	System        string    `json:"system,omitempty"`
+	Messages      []Message `json:"messages"`
+	MaxTokens     int       `json:"max_tokens"`
+	StopSequences []string  `json:"stop_sequences,omitempty"`
+	Temperature   float64   `json:"temperature,omitempty"`
+	TopP          float64   `json:"top_p,omitempty"`
+	TopK          int       `json:"top_k,omitempty"`
 	//ClaudeMetadata    `json:"metadata,omitempty"`
 	Stream bool `json:"stream,omitempty"`
 }

-type ClaudeResponseError struct {
-	Error ClaudeError `json:"error,omitempty"`
+type Usage struct {
+	InputTokens  int `json:"input_tokens"`
+	OutputTokens int `json:"output_tokens,omitempty"`
 }
 type ClaudeResponse struct {
-	Completion string       `json:"completion"`
-	StopReason string       `json:"stop_reason"`
-	Model      string       `json:"model"`
-	Usage      *types.Usage `json:"usage,omitempty"`
-	ClaudeResponseError
+	Content      []ResContent `json:"content"`
+	Id           string       `json:"id"`
+	Role         string       `json:"role"`
+	StopReason   string       `json:"stop_reason"`
+	StopSequence string       `json:"stop_sequence,omitempty"`
+	Model        string       `json:"model"`
+	Usage        `json:"usage,omitempty"`
+	Error        ClaudeError `json:"error,omitempty"`
+}
+
+type Delta struct {
+	Type         string `json:"type,omitempty"`
+	Text         string `json:"text,omitempty"`
+	StopReason   string `json:"stop_reason,omitempty"`
+	StopSequence string `json:"stop_sequence,omitempty"`
+}
+
+type ClaudeStreamResponse struct {
+	Type    string         `json:"type"`
+	Message ClaudeResponse `json:"message,omitempty"`
+	Index   int            `json:"index,omitempty"`
+	Delta   Delta          `json:"delta,omitempty"`
+	Usage   Usage          `json:"usage,omitempty"`
+	Error   ClaudeError    `json:"error,omitempty"`
 }
--- a/web/src/views/Channel/type/Config.js
+++ b/web/src/views/Channel/type/Config.js
@ -59,8 +59,8 @@ const typeConfig = {
  },
  14: {
    input: {
-      models: ['claude-instant-1', 'claude-2', 'claude-2.0', 'claude-2.1'],
-      test_model: 'claude-2'
+      models: ['claude-instant-1.2', 'claude-2.0', 'claude-2.1','claude-3-opus-20240229','claude-3-sonnet-20240229'],
+      test_model: 'claude-3-sonnet-20240229'
    },
    modelGroup: 'Anthropic'
  },