feat: implement thinking budget control in model name

2025-06-15 21:12:56 +08:00
parent d2b47969da
commit e735377218
5 changed files with 69 additions and 9 deletions
--- a/relay/channel/gemini/adaptor.go
+++ b/relay/channel/gemini/adaptor.go
@@ -72,8 +72,11 @@ 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") {
+		// 新增逻辑：处理 -thinking-<budget> 格式
+		if strings.Contains(info.OriginModelName, "-thinking-") {
+			parts := strings.Split(info.UpstreamModelName, "-thinking-")
+			info.UpstreamModelName = parts[0]
+		} else if strings.HasSuffix(info.OriginModelName, "-thinking") { // 旧的适配
 			info.UpstreamModelName = strings.TrimSuffix(info.UpstreamModelName, "-thinking")
 		} else if strings.HasSuffix(info.OriginModelName, "-nothinking") {
 			info.UpstreamModelName = strings.TrimSuffix(info.UpstreamModelName, "-nothinking")
--- a/relay/channel/gemini/relay-gemini.go
+++ b/relay/channel/gemini/relay-gemini.go
@@ -12,6 +12,7 @@ import (
 	"one-api/relay/helper"
 	"one-api/service"
 	"one-api/setting/model_setting"
+	"strconv"
 	"strings"
 	"unicode/utf8"

@@ -36,6 +37,13 @@ var geminiSupportedMimeTypes = map[string]bool{
 	"video/flv":       true,
 }

+// Gemini 允许的思考预算范围
+const (
+	pro25MinBudget   = 128
+	pro25MaxBudget   = 32768
+	flash25MaxBudget = 24576
+)
+
 // Setting safety to the lowest possible values since Gemini is already powerless enough
 func CovertGemini2OpenAI(textRequest dto.GeneralOpenAIRequest, info *relaycommon.RelayInfo) (*GeminiChatRequest, error) {

@@ -57,7 +65,40 @@ func CovertGemini2OpenAI(textRequest dto.GeneralOpenAIRequest, info *relaycommon
 	}

 	if model_setting.GetGeminiSettings().ThinkingAdapterEnabled {
-		if strings.HasSuffix(info.OriginModelName, "-thinking") {
+		// 新增逻辑：处理 -thinking-<budget> 格式
+		if strings.Contains(info.OriginModelName, "-thinking-") {
+			parts := strings.SplitN(info.OriginModelName, "-thinking-", 2)
+			if len(parts) == 2 && parts[1] != "" {
+				if budgetTokens, err := strconv.Atoi(parts[1]); err == nil {
+					// 从模型名称成功解析预算
+					isNew25Pro := strings.HasPrefix(info.OriginModelName, "gemini-2.5-pro") &&
+						!strings.HasPrefix(info.OriginModelName, "gemini-2.5-pro-preview-05-06") &&
+						!strings.HasPrefix(info.OriginModelName, "gemini-2.5-pro-preview-03-25")
+
+					if isNew25Pro {
+						// 新的2.5pro模型：ThinkingBudget范围为128-32768
+						if budgetTokens < pro25MinBudget {
+							budgetTokens = pro25MinBudget
+						} else if budgetTokens > pro25MaxBudget {
+							budgetTokens = pro25MaxBudget
+						}
+					} else {
+						// 其他模型：ThinkingBudget范围为0-24576
+						if budgetTokens < 0 {
+							budgetTokens = 0
+						} else if budgetTokens > flash25MaxBudget {
+							budgetTokens = flash25MaxBudget
+						}
+					}
+
+					geminiRequest.GenerationConfig.ThinkingConfig = &GeminiThinkingConfig{
+						ThinkingBudget:  common.GetPointer(budgetTokens),
+						IncludeThoughts: true,
+					}
+				}
+				// 如果解析失败，则不设置ThinkingConfig，静默处理
+			}
+		} else if strings.HasSuffix(info.OriginModelName, "-thinking") { // 保留旧逻辑以兼容
 			// 硬编码不支持 ThinkingBudget 的旧模型
 			unsupportedModels := []string{
 				"gemini-2.5-pro-preview-05-06",
--- a/setting/operation_setting/model-ratio.go
+++ b/setting/operation_setting/model-ratio.go
@@ -142,6 +142,11 @@ var defaultModelRatio = map[string]float64{
 	"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,
+	"gemini-2.5-flash-preview-05-20":            0.075,
+	"gemini-2.5-flash-preview-05-20-thinking":   0.075,
+	"gemini-2.5-flash-preview-05-20-nothinking": 0.075,
+	"gemini-2.5-flash-thinking-*":               0.075, // 用于为后续所有2.5 flash thinking budget 模型设置默认倍率
+	"gemini-2.5-pro-thinking-*":                 0.625, // 用于为后续所有2.5 pro thinking budget 模型设置默认倍率
 	"text-embedding-004":                        0.001,
 	"chatglm_turbo":                             0.3572,     // ￥0.005 / 1k tokens
 	"chatglm_pro":                               0.7143,     // ￥0.01 / 1k tokens
@@ -345,7 +350,14 @@ func UpdateModelRatioByJSONString(jsonStr string) error {
 func GetModelRatio(name string) (float64, bool) {
 	modelRatioMapMutex.RLock()
 	defer modelRatioMapMutex.RUnlock()
-
+	// 处理带有思考预算的模型名称，方便统一定价
+	handleThinkingBudgetModel := func(prefix, wildcard string) {
+		if strings.HasPrefix(name, prefix) && strings.Contains(name, "-thinking-") {
+			name = wildcard
+		}
+	}
+	handleThinkingBudgetModel("gemini-2.5-flash", "gemini-2.5-flash-thinking-*")
+	handleThinkingBudgetModel("gemini-2.5-pro", "gemini-2.5-pro-thinking-*")
 	if strings.HasPrefix(name, "gpt-4-gizmo") {
 		name = "gpt-4-gizmo-*"
 	}
@@ -470,9 +482,9 @@ func getHardcodedCompletionModelRatio(name string) (float64, bool) {
 			return 4, true
 		} else if strings.HasPrefix(name, "gemini-2.0") {
 			return 4, true
-		} else if strings.HasPrefix(name, "gemini-2.5-pro-preview") {
+		} else if strings.HasPrefix(name, "gemini-2.5-pro") { // 移除preview来增加兼容性，这里假设正式版的倍率和preview一致
 			return 8, true
-		} else if strings.HasPrefix(name, "gemini-2.5-flash-preview") {
+		} else if strings.HasPrefix(name, "gemini-2.5-flash") { // 同上
 			if strings.HasSuffix(name, "-nothinking") {
 				return 4, false
 			} else {
--- a/web/src/i18n/locales/en.json
+++ b/web/src/i18n/locales/en.json
@@ -1373,6 +1373,9 @@
  "示例": "Example",
  "缺省 MaxTokens": "Default MaxTokens",
  "启用Claude思考适配（-thinking后缀）": "Enable Claude thinking adaptation (-thinking suffix)",
+  "启用Gemini思考后缀适配": "Enable Gemini thinking suffix adaptation",
+  "适配-thinking、-thinking-预算数字和-nothinking后缀": "Adapt -thinking, -thinking-budgetNumber, and -nothinking suffixes",
+  "思考预算占比": "Thinking budget ratio",
  "Claude思考适配 BudgetTokens = MaxTokens * BudgetTokens 百分比": "Claude thinking adaptation BudgetTokens = MaxTokens * BudgetTokens percentage",
  "思考适配 BudgetTokens 百分比": "Thinking adaptation BudgetTokens percentage",
  "0.1-1之间的小数": "Decimal between 0.1 and 1",
--- a/web/src/pages/Setting/Model/SettingGeminiModel.js
+++ b/web/src/pages/Setting/Model/SettingGeminiModel.js
@@ -173,7 +173,8 @@ export default function SettingGeminiModel(props) {
                <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，" +
-                    "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"
+                    "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。" +
+                    "支持使用 gemini-2.5-pro-preview-06-05-thinking-128 格式来精确传递思考预算。"
                  )}
                </Text>
              </Col>
@@ -183,7 +184,7 @@ export default function SettingGeminiModel(props) {
                <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"}
+                  extraText={t('适配 -thinking、-thinking-预算数字 和 -nothinking 后缀')}
                  onChange={(value) =>
                    setInputs({
                      ...inputs,
@@ -205,7 +206,7 @@ export default function SettingGeminiModel(props) {
            <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）')}
+                  label={t('思考预算占比')}
                  field={'gemini.thinking_adapter_budget_tokens_percentage'}
                  initValue={''}
                  extraText={t('0.1-1之间的小数')}