xinyin025/new-api
1
0
Fork 0
mirror of https://github.com/linux-do/new-api.git synced 2025-09-17 16:06:38 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

fix: fix embedding

Browse Source
This commit is contained in:
CaIon 2024-03-21 17:39:05 +08:00
parent 031957714a
commit c4b3d3a975
6 changed files with 60 additions and 37 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
dto/text_response.go
View File

@ -1,12 +1,16 @@
package dto package dto
type TextResponseWithError struct { type TextResponseWithError struct {
Choices []OpenAITextResponseChoice `json:"choices"` Choices []OpenAITextResponseChoice `json:"choices"`
Object string `json:"object"`
Data []OpenAIEmbeddingResponseItem `json:"data"`
Model string `json:"model"`
Usage `json:"usage"` Usage `json:"usage"`
Error OpenAIError `json:"error"` Error OpenAIError `json:"error"`
} }
type TextResponse struct { type TextResponse struct {
Model string `json:"model"`
Choices []OpenAITextResponseChoice `json:"choices"` Choices []OpenAITextResponseChoice `json:"choices"`
Usage `json:"usage"` Usage `json:"usage"`
} }

2
relay/channel/ollama/adaptor.go
View File

@ -45,7 +45,7 @@ func (a *Adaptor) DoResponse(c *gin.Context, resp *http.Response, info *relaycom
err, responseText = openai.OpenaiStreamHandler(c, resp, info.RelayMode) err, responseText = openai.OpenaiStreamHandler(c, resp, info.RelayMode)
usage, _ = service.ResponseText2Usage(responseText, info.UpstreamModelName, info.PromptTokens) usage, _ = service.ResponseText2Usage(responseText, info.UpstreamModelName, info.PromptTokens)
} else { } else {
err, usage, sensitiveResp = openai.OpenaiHandler(c, resp, info.PromptTokens, info.UpstreamModelName) err, usage, sensitiveResp = openai.OpenaiHandler(c, resp, info.PromptTokens, info.UpstreamModelName, info.RelayMode)
} }
return return
} }

2
relay/channel/openai/adaptor.go
View File

@ -77,7 +77,7 @@ func (a *Adaptor) DoResponse(c *gin.Context, resp *http.Response, info *relaycom
err, responseText = OpenaiStreamHandler(c, resp, info.RelayMode) err, responseText = OpenaiStreamHandler(c, resp, info.RelayMode)
usage, _ = service.ResponseText2Usage(responseText, info.UpstreamModelName, info.PromptTokens) usage, _ = service.ResponseText2Usage(responseText, info.UpstreamModelName, info.PromptTokens)
} else { } else {
err, usage, sensitiveResp = OpenaiHandler(c, resp, info.PromptTokens, info.UpstreamModelName) err, usage, sensitiveResp = OpenaiHandler(c, resp, info.PromptTokens, info.UpstreamModelName, info.RelayMode)
} }
return return
} }

83
relay/channel/openai/relay-openai.go
View File

@ -124,8 +124,8 @@ func OpenaiStreamHandler(c *gin.Context, resp *http.Response, relayMode int) (*d
return nil, responseTextBuilder.String() return nil, responseTextBuilder.String()
} }
func OpenaiHandler(c *gin.Context, resp *http.Response, promptTokens int, model string) (*dto.OpenAIErrorWithStatusCode, *dto.Usage, *dto.SensitiveResponse) { func OpenaiHandler(c *gin.Context, resp *http.Response, promptTokens int, model string, relayMode int) (*dto.OpenAIErrorWithStatusCode, *dto.Usage, *dto.SensitiveResponse) {
var textResponseWithError dto.TextResponseWithError var responseWithError dto.TextResponseWithError
responseBody, err := io.ReadAll(resp.Body) responseBody, err := io.ReadAll(resp.Body)
if err != nil { if err != nil {
return service.OpenAIErrorWrapper(err, "read_response_body_failed", http.StatusInternalServerError), nil, nil return service.OpenAIErrorWrapper(err, "read_response_body_failed", http.StatusInternalServerError), nil, nil
@ -134,62 +134,81 @@ func OpenaiHandler(c *gin.Context, resp *http.Response, promptTokens int, model
if err != nil { if err != nil {
return service.OpenAIErrorWrapper(err, "close_response_body_failed", http.StatusInternalServerError), nil, nil return service.OpenAIErrorWrapper(err, "close_response_body_failed", http.StatusInternalServerError), nil, nil
} }
err = json.Unmarshal(responseBody, &textResponseWithError) err = json.Unmarshal(responseBody, &responseWithError)
if err != nil { if err != nil {
log.Printf("unmarshal_response_body_failed: body: %s, err: %v", string(responseBody), err) log.Printf("unmarshal_response_body_failed: body: %s, err: %v", string(responseBody), err)
return service.OpenAIErrorWrapper(err, "unmarshal_response_body_failed", http.StatusInternalServerError), nil, nil return service.OpenAIErrorWrapper(err, "unmarshal_response_body_failed", http.StatusInternalServerError), nil, nil
} }
if textResponseWithError.Error.Type != "" { if responseWithError.Error.Type != "" {
return &dto.OpenAIErrorWithStatusCode{ return &dto.OpenAIErrorWithStatusCode{
Error: textResponseWithError.Error, Error: responseWithError.Error,
StatusCode: resp.StatusCode, StatusCode: resp.StatusCode,
}, nil, nil }, nil, nil
} }
textResponse := &dto.TextResponse{
Choices: textResponseWithError.Choices,
Usage: textResponseWithError.Usage,
}
checkSensitive := constant.ShouldCheckCompletionSensitive() checkSensitive := constant.ShouldCheckCompletionSensitive()
sensitiveWords := make([]string, 0) sensitiveWords := make([]string, 0)
triggerSensitive := false triggerSensitive := false
if textResponse.Usage.TotalTokens == 0 || checkSensitive { usage := &responseWithError.Usage
completionTokens := 0
for i, choice := range textResponse.Choices { //textResponse := &dto.TextResponse{
stringContent := string(choice.Message.Content) // Choices: responseWithError.Choices,
ctkm, _, _ := service.CountTokenText(stringContent, model, false) // Usage: responseWithError.Usage,
completionTokens += ctkm //}
if checkSensitive { var doResponseBody []byte
sensitive, words, stringContent := service.SensitiveWordReplace(stringContent, false)
if sensitive { switch relayMode {
triggerSensitive = true case relayconstant.RelayModeEmbeddings:
msg := choice.Message embeddingResponse := &dto.OpenAIEmbeddingResponse{
msg.Content = common.StringToByteSlice(stringContent) Object: responseWithError.Object,
textResponse.Choices[i].Message = msg Data: responseWithError.Data,
sensitiveWords = append(sensitiveWords, words...) Model: responseWithError.Model,
Usage: *usage,
}
doResponseBody, err = json.Marshal(embeddingResponse)
default:
if responseWithError.Usage.TotalTokens == 0 || checkSensitive {
completionTokens := 0
for i, choice := range responseWithError.Choices {
stringContent := string(choice.Message.Content)
ctkm, _, _ := service.CountTokenText(stringContent, model, false)
completionTokens += ctkm
if checkSensitive {
sensitive, words, stringContent := service.SensitiveWordReplace(stringContent, false)
if sensitive {
triggerSensitive = true
msg := choice.Message
msg.Content = common.StringToByteSlice(stringContent)
responseWithError.Choices[i].Message = msg
sensitiveWords = append(sensitiveWords, words...)
}
} }
} }
responseWithError.Usage = dto.Usage{
PromptTokens: promptTokens,
CompletionTokens: completionTokens,
TotalTokens: promptTokens + completionTokens,
}
} }
textResponse.Usage = dto.Usage{ textResponse := &dto.TextResponse{
PromptTokens: promptTokens, Choices: responseWithError.Choices,
CompletionTokens: completionTokens, Model: responseWithError.Model,
TotalTokens: promptTokens + completionTokens, Usage: *usage,
} }
doResponseBody, err = json.Marshal(textResponse)
} }
if checkSensitive && triggerSensitive && constant.StopOnSensitiveEnabled { if checkSensitive && triggerSensitive && constant.StopOnSensitiveEnabled {
sensitiveWords = common.RemoveDuplicate(sensitiveWords) sensitiveWords = common.RemoveDuplicate(sensitiveWords)
return service.OpenAIErrorWrapper(errors.New(fmt.Sprintf("sensitive words detected on response: %s", return service.OpenAIErrorWrapper(errors.New(fmt.Sprintf("sensitive words detected on response: %s",
strings.Join(sensitiveWords, ", "))), "sensitive_words_detected", http.StatusBadRequest), strings.Join(sensitiveWords, ", "))), "sensitive_words_detected", http.StatusBadRequest),
&textResponse.Usage, &dto.SensitiveResponse{ usage, &dto.SensitiveResponse{
SensitiveWords: sensitiveWords, SensitiveWords: sensitiveWords,
} }
} else { } else {
responseBody, err = json.Marshal(textResponse)
// Reset response body // Reset response body
resp.Body = io.NopCloser(bytes.NewBuffer(responseBody)) resp.Body = io.NopCloser(bytes.NewBuffer(doResponseBody))
// We shouldn't set the header before we parse the response body, because the parse part may fail. // 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. // 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. // So the httpClient will be confused by the response.
@ -207,5 +226,5 @@ func OpenaiHandler(c *gin.Context, resp *http.Response, promptTokens int, model
return service.OpenAIErrorWrapper(err, "close_response_body_failed", http.StatusInternalServerError), nil, nil return service.OpenAIErrorWrapper(err, "close_response_body_failed", http.StatusInternalServerError), nil, nil
} }
} }
return nil, &textResponse.Usage, nil return nil, usage, nil
} }

2
relay/channel/perplexity/adaptor.go
View File

@ -49,7 +49,7 @@ func (a *Adaptor) DoResponse(c *gin.Context, resp *http.Response, info *relaycom
err, responseText = openai.OpenaiStreamHandler(c, resp, info.RelayMode) err, responseText = openai.OpenaiStreamHandler(c, resp, info.RelayMode)
usage, _ = service.ResponseText2Usage(responseText, info.UpstreamModelName, info.PromptTokens) usage, _ = service.ResponseText2Usage(responseText, info.UpstreamModelName, info.PromptTokens)
} else { } else {
err, usage, sensitiveResp = openai.OpenaiHandler(c, resp, info.PromptTokens, info.UpstreamModelName) err, usage, sensitiveResp = openai.OpenaiHandler(c, resp, info.PromptTokens, info.UpstreamModelName, info.RelayMode)
} }
return return
} }

2
relay/channel/zhipu_4v/adaptor.go
View File

@ -50,7 +50,7 @@ func (a *Adaptor) DoResponse(c *gin.Context, resp *http.Response, info *relaycom
err, responseText = openai.OpenaiStreamHandler(c, resp, info.RelayMode) err, responseText = openai.OpenaiStreamHandler(c, resp, info.RelayMode)
usage, _ = service.ResponseText2Usage(responseText, info.UpstreamModelName, info.PromptTokens) usage, _ = service.ResponseText2Usage(responseText, info.UpstreamModelName, info.PromptTokens)
} else { } else {
err, usage, sensitiveResp = openai.OpenaiHandler(c, resp, info.PromptTokens, info.UpstreamModelName) err, usage, sensitiveResp = openai.OpenaiHandler(c, resp, info.PromptTokens, info.UpstreamModelName, info.RelayMode)
} }
return return
} }