feat: add gemini thinking suffix support #981

dto/dalle.go

@@ -1,5 +1,7 @@
 package dto
 
+import "encoding/json"
+
 type ImageRequest struct {
 	Model          string          `json:"model"`
 	Prompt         string          `json:"prompt" binding:"required"`
@@ -9,6 +11,7 @@ type ImageRequest struct {
 	ResponseFormat string          `json:"response_format,omitempty"`
 	Style          string          `json:"style,omitempty"`
 	User           string          `json:"user,omitempty"`
+	ExtraFields    json.RawMessage `json:"extra_fields,omitempty"`
 }
 
 type ImageResponse struct {

dto/openai_response.go

@@ -173,3 +173,17 @@ type Usage struct {
 	PromptTokensDetails    InputTokenDetails  `json:"prompt_tokens_details"`
 	CompletionTokenDetails OutputTokenDetails `json:"completion_tokens_details"`
 }
+
+type InputTokenDetails struct {
+	CachedTokens         int `json:"cached_tokens"`
+	CachedCreationTokens int `json:"-"`
+	TextTokens           int `json:"text_tokens"`
+	AudioTokens          int `json:"audio_tokens"`
+	ImageTokens          int `json:"image_tokens"`
+}
+
+type OutputTokenDetails struct {
+	TextTokens      int `json:"text_tokens"`
+	AudioTokens     int `json:"audio_tokens"`
+	ReasoningTokens int `json:"reasoning_tokens"`
+}

dto/realtime.go

@@ -43,20 +43,6 @@ type RealtimeUsage struct {
 	OutputTokenDetails OutputTokenDetails `json:"output_token_details"`
 }
 
-type InputTokenDetails struct {
-	CachedTokens         int `json:"cached_tokens"`
-	CachedCreationTokens int `json:"-"`
-	TextTokens           int `json:"text_tokens"`
-	AudioTokens          int `json:"audio_tokens"`
-	ImageTokens          int `json:"image_tokens"`
-}
-
-type OutputTokenDetails struct {
-	TextTokens      int `json:"text_tokens"`
-	AudioTokens     int `json:"audio_tokens"`
-	ReasoningTokens int `json:"reasoning_tokens"`
-}
-
 type RealtimeSession struct {
 	Modalities              []string                `json:"modalities"`
 	Instructions            string                  `json:"instructions"`

relay/channel/gemini/adaptor.go

@@ -12,7 +12,6 @@ import (
 	relaycommon "one-api/relay/common"
 	"one-api/service"
 	"one-api/setting/model_setting"
-
 	"strings"
 
 	"github.com/gin-gonic/gin"
@@ -70,6 +69,16 @@ func (a *Adaptor) Init(info *relaycommon.RelayInfo) {
 }
 
 func (a *Adaptor) GetRequestURL(info *relaycommon.RelayInfo) (string, error) {
+
+	if model_setting.GetGeminiSettings().ThinkingAdapterEnabled {
+		// suffix -thinking and -nothinking
+		if strings.HasSuffix(info.OriginModelName, "-thinking") {
+			info.UpstreamModelName = strings.TrimSuffix(info.UpstreamModelName, "-thinking")
+		} else if strings.HasSuffix(info.OriginModelName, "-nothinking") {
+			info.UpstreamModelName = strings.TrimSuffix(info.UpstreamModelName, "-nothinking")
+		}
+	}
+
 	version := model_setting.GetGeminiVersionSetting(info.UpstreamModelName)
 
 	if strings.HasPrefix(info.UpstreamModelName, "imagen") {
@@ -99,11 +108,13 @@ func (a *Adaptor) ConvertOpenAIRequest(c *gin.Context, info *relaycommon.RelayIn
 	if request == nil {
 		return nil, errors.New("request is nil")
 	}
-	ai, err := CovertGemini2OpenAI(*request, info)
+
+	geminiRequest, err := CovertGemini2OpenAI(*request, info)
 	if err != nil {
 		return nil, err
 	}
-	return ai, nil
+
+	return geminiRequest, nil
 }
 
 func (a *Adaptor) ConvertRerankRequest(c *gin.Context, relayMode int, request dto.RerankRequest) (any, error) {
@@ -165,6 +176,18 @@ func (a *Adaptor) DoResponse(c *gin.Context, resp *http.Response, info *relaycom
 	} else {
 		err, usage = GeminiChatHandler(c, resp, info)
 	}
+
+	//if usage.(*dto.Usage).CompletionTokenDetails.ReasoningTokens > 100 {
+	//	// 没有请求-thinking的情况下，产生思考token，则按照思考模型计费
+	//	if !strings.HasSuffix(info.OriginModelName, "-thinking") &&
+	//		!strings.HasSuffix(info.OriginModelName, "-nothinking") {
+	//		thinkingModelName := info.OriginModelName + "-thinking"
+	//		if operation_setting.SelfUseModeEnabled || helper.ContainPriceOrRatio(thinkingModelName) {
+	//			info.OriginModelName = thinkingModelName
+	//		}
+	//	}
+	//}
+
 	return
 }
 

relay/channel/gemini/dto.go

@@ -8,6 +8,15 @@ type GeminiChatRequest struct {
 	SystemInstructions *GeminiChatContent         `json:"system_instruction,omitempty"`
 }
 
+type GeminiThinkingConfig struct {
+	IncludeThoughts bool `json:"includeThoughts,omitempty"`
+	ThinkingBudget  *int `json:"thinkingBudget,omitempty"`
+}
+
+func (c *GeminiThinkingConfig) SetThinkingBudget(budget int) {
+	c.ThinkingBudget = &budget
+}
+
 type GeminiInlineData struct {
 	MimeType string `json:"mimeType"`
 	Data     string `json:"data"`
@@ -81,6 +90,7 @@ type GeminiChatGenerationConfig struct {
 	ResponseSchema     any                   `json:"responseSchema,omitempty"`
 	Seed               int64                 `json:"seed,omitempty"`
 	ResponseModalities []string              `json:"responseModalities,omitempty"`
+	ThinkingConfig     *GeminiThinkingConfig `json:"thinkingConfig,omitempty"`
 }
 
 type GeminiChatCandidate struct {
@@ -109,6 +119,7 @@ type GeminiUsageMetadata struct {
 	PromptTokenCount     int `json:"promptTokenCount"`
 	CandidatesTokenCount int `json:"candidatesTokenCount"`
 	TotalTokenCount      int `json:"totalTokenCount"`
+	ThoughtsTokenCount   int `json:"thoughtsTokenCount"`
 }
 
 // Imagen related structs

relay/channel/gemini/relay-gemini.go

@@ -23,12 +23,14 @@ func CovertGemini2OpenAI(textRequest dto.GeneralOpenAIRequest, info *relaycommon
 
 	geminiRequest := GeminiChatRequest{
 		Contents: make([]GeminiChatContent, 0, len(textRequest.Messages)),
-		//SafetySettings: []GeminiChatSafetySettings{},
 		GenerationConfig: GeminiChatGenerationConfig{
 			Temperature:     textRequest.Temperature,
 			TopP:            textRequest.TopP,
 			MaxOutputTokens: textRequest.MaxTokens,
 			Seed:            int64(textRequest.Seed),
+			ThinkingConfig: &GeminiThinkingConfig{
+				IncludeThoughts: true,
+			},
 		},
 	}
 
@@ -39,6 +41,18 @@ func CovertGemini2OpenAI(textRequest dto.GeneralOpenAIRequest, info *relaycommon
 		}
 	}
 
+	if model_setting.GetGeminiSettings().ThinkingAdapterEnabled {
+		if strings.HasSuffix(info.OriginModelName, "-thinking") {
+			budgetTokens := model_setting.GetGeminiSettings().ThinkingAdapterBudgetTokensPercentage * float64(geminiRequest.GenerationConfig.MaxOutputTokens)
+			if budgetTokens == 0 || budgetTokens > 24576 {
+				budgetTokens = 24576
+			}
+			geminiRequest.GenerationConfig.ThinkingConfig.SetThinkingBudget(int(budgetTokens))
+		} else if strings.HasSuffix(info.OriginModelName, "-nothinking") {
+			geminiRequest.GenerationConfig.ThinkingConfig.SetThinkingBudget(0)
+		}
+	}
+
 	safetySettings := make([]GeminiChatSafetySettings, 0, len(SafetySettingList))
 	for _, category := range SafetySettingList {
 		safetySettings = append(safetySettings, GeminiChatSafetySettings{
@@ -644,6 +658,7 @@ func GeminiChatStreamHandler(c *gin.Context, resp *http.Response, info *relaycom
 		if geminiResponse.UsageMetadata.TotalTokenCount != 0 {
 			usage.PromptTokens = geminiResponse.UsageMetadata.PromptTokenCount
 			usage.CompletionTokens = geminiResponse.UsageMetadata.CandidatesTokenCount
+			usage.CompletionTokenDetails.ReasoningTokens = geminiResponse.UsageMetadata.ThoughtsTokenCount
 		}
 		err = helper.ObjectData(c, response)
 		if err != nil {
@@ -666,7 +681,7 @@ func GeminiChatStreamHandler(c *gin.Context, resp *http.Response, info *relaycom
 
 	usage.TotalTokens = usage.PromptTokens + usage.CompletionTokens
 	usage.PromptTokensDetails.TextTokens = usage.PromptTokens
-	usage.CompletionTokenDetails.TextTokens = usage.CompletionTokens
+	//usage.CompletionTokenDetails.TextTokens = usage.CompletionTokens
 
 	if info.ShouldIncludeUsage {
 		response = helper.GenerateFinalUsageResponse(id, createAt, info.UpstreamModelName, *usage)
@@ -712,6 +727,9 @@ func GeminiChatHandler(c *gin.Context, resp *http.Response, info *relaycommon.Re
 		CompletionTokens: geminiResponse.UsageMetadata.CandidatesTokenCount,
 		TotalTokens:      geminiResponse.UsageMetadata.TotalTokenCount,
 	}
+
+	usage.CompletionTokenDetails.ReasoningTokens = geminiResponse.UsageMetadata.ThoughtsTokenCount
+
 	fullTextResponse.Usage = usage
 	jsonResponse, err := json.Marshal(fullTextResponse)
 	if err != nil {

relay/helper/price.go

@@ -49,11 +49,7 @@ func ModelPriceHelper(c *gin.Context, info *relaycommon.RelayInfo, promptTokens
 				}
 			}
 			if !acceptUnsetRatio {
-				if info.UserId == 1 {
+				return PriceData{}, fmt.Errorf("模型 %s 倍率或价格未配置，请联系管理员设置或开始自用模式；Model %s ratio or price not set, please set or start self-use mode", info.OriginModelName, info.OriginModelName)
-					return PriceData{}, fmt.Errorf("模型 %s 倍率或价格未配置，请设置或开始自用模式；Model %s ratio or price not set, please set or start self-use mode", info.OriginModelName, info.OriginModelName)
-				} else {
-					return PriceData{}, fmt.Errorf("模型 %s 倍率或价格未配置, 请联系管理员设置；Model %s ratio or price not set, please contact administrator to set", info.OriginModelName, info.OriginModelName)
-				}
 			}
 		}
 		completionRatio = operation_setting.GetCompletionRatio(info.OriginModelName)
@@ -82,3 +78,15 @@ func ModelPriceHelper(c *gin.Context, info *relaycommon.RelayInfo, promptTokens
 
 	return priceData, nil
 }
+
+func ContainPriceOrRatio(modelName string) bool {
+	_, ok := operation_setting.GetModelPrice(modelName, false)
+	if ok {
+		return true
+	}
+	_, ok = operation_setting.GetModelRatio(modelName)
+	if ok {
+		return true
+	}
+	return false
+}

setting/model_setting/gemini.go

@@ -9,6 +9,8 @@ type GeminiSettings struct {
 	SafetySettings                        map[string]string `json:"safety_settings"`
 	VersionSettings                       map[string]string `json:"version_settings"`
 	SupportedImagineModels                []string          `json:"supported_imagine_models"`
+	ThinkingAdapterEnabled                bool              `json:"thinking_adapter_enabled"`
+	ThinkingAdapterBudgetTokensPercentage float64           `json:"thinking_adapter_budget_tokens_percentage"`
 }
 
 // 默认配置
@@ -25,6 +27,8 @@ var defaultGeminiSettings = GeminiSettings{
 		"gemini-2.0-flash-exp-image-generation",
 		"gemini-2.0-flash-exp",
 	},
+	ThinkingAdapterEnabled:                false,
+	ThinkingAdapterBudgetTokensPercentage: 0.6,
 }
 
 // 全局实例

setting/operation_setting/model-ratio.go

@@ -136,6 +136,9 @@ var defaultModelRatio = map[string]float64{
 	"gemini-2.0-flash":                          0.05,
 	"gemini-2.5-pro-exp-03-25":                  0.625,
 	"gemini-2.5-pro-preview-03-25":              0.625,
+	"gemini-2.5-flash-preview-04-17":            0.075,
+	"gemini-2.5-flash-preview-04-17-thinking":   0.075,
+	"gemini-2.5-flash-preview-04-17-nothinking": 0.075,
 	"text-embedding-004":                        0.001,
 	"chatglm_turbo":                             0.3572,     // ￥0.005 / 1k tokens
 	"chatglm_pro":                               0.7143,     // ￥0.01 / 1k tokens
@@ -275,8 +278,6 @@ func InitModelSettings() {
 	cacheRatioMapMutex.Lock()
 	cacheRatioMap = defaultCacheRatio
 	cacheRatioMapMutex.Unlock()
-
-	common.SysLog("model settings initialized")
 }
 
 func GetModelPriceMap() map[string]float64 {
@@ -459,6 +460,12 @@ func getHardcodedCompletionModelRatio(name string) (float64, bool) {
 			return 4, true
 		} else if strings.HasPrefix(name, "gemini-2.5-pro-preview") {
 			return 8, true
+		} else if strings.HasPrefix(name, "gemini-2.5-flash-preview") {
+			if strings.HasSuffix(name, "-nothinking") {
+				return 4, false
+			} else {
+				return 3.5 / 0.6, false
+			}
 		}
 		return 4, false
 	}

web/src/components/ModelSetting.js

@@ -20,6 +20,8 @@ const ModelSetting = () => {
     'global.pass_through_request_enabled': false,
     'general_setting.ping_interval_enabled': false,
     'general_setting.ping_interval_seconds': 60,
+    'gemini.thinking_adapter_enabled': false,
+    'gemini.thinking_adapter_budget_tokens_percentage': 0.6,
   });
 
   let [loading, setLoading] = useState(false);

web/src/pages/Setting/Model/SettingGeminiModel.js

@@ -9,6 +9,7 @@ import {
   verifyJSON,
 } from '../../../helpers';
 import { useTranslation } from 'react-i18next';
+import Text from '@douyinfe/semi-ui/lib/es/typography/text.js';
 
 const GEMINI_SETTING_EXAMPLE = {
   default: 'OFF',
@@ -27,6 +28,8 @@ export default function SettingGeminiModel(props) {
     'gemini.safety_settings': '',
     'gemini.version_settings': '',
     'gemini.supported_imagine_models': [],
+    'gemini.thinking_adapter_enabled': false,
+    'gemini.thinking_adapter_budget_tokens_percentage': 0.6,
   });
   const refForm = useRef();
   const [inputsRow, setInputsRow] = useState(inputs);
@@ -151,13 +154,69 @@ export default function SettingGeminiModel(props) {
                 />
               </Col>
             </Row>
+          </Form.Section>
+
+          <Form.Section text={t('Gemini思考适配设置')}>
+            <Row>
+              <Col span={16}>
+                <Text>
+                  {t(
+                    "和Claude不同，默认情况下Gemini的思考模型会自动决定要不要思考，就算不开启适配模型也可以正常使用，" +
+                    "-nothinking后缀（BudgetTokens=0，思考关闭）也会返回少量的思考token，这是gemini的特性，" +
+                    "如果您需要计费，推荐设置无后缀模型价格按思考价格设置"
+                  )}
+                </Text>
+              </Col>
+            </Row>
+            <Row>
+              <Col span={16}>
+                <Form.Switch
+                  label={t('启用Gemini思考后缀适配')}
+                  field={'gemini.thinking_adapter_enabled'}
+                  extraText={"适配-thinking和-nothinking后缀"}
+                  onChange={(value) =>
+                    setInputs({
+                      ...inputs,
+                      'gemini.thinking_adapter_enabled': value,
+                    })
+                  }
+                />
+              </Col>
+            </Row>
+            <Row>
+              <Col span={16}>
+                <Text>
+                  {t(
+                    'Gemini思考适配 BudgetTokens = MaxTokens * BudgetTokens 百分比',
+                  )}
+                </Text>
+              </Col>
+            </Row>
+            <Row>
+              <Col xs={24} sm={12} md={8} lg={8} xl={8}>
+                <Form.InputNumber
+                  label={t('请求模型带-thinking后缀的BudgetTokens数（超出24576的部分将被忽略）')}
+                  field={'gemini.thinking_adapter_budget_tokens_percentage'}
+                  initValue={''}
+                  extraText={t('0.1-1之间的小数')}
+                  min={0.1}
+                  max={1}
+                  onChange={(value) =>
+                    setInputs({
+                      ...inputs,
+                      'gemini.thinking_adapter_budget_tokens_percentage': value,
+                    })
+                  }
+                />
+              </Col>
+            </Row>
+          </Form.Section>
 
           <Row>
             <Button size='default' onClick={onSubmit}>
               {t('保存')}
             </Button>
           </Row>
-          </Form.Section>
         </Form>
       </Spin>
     </>