Merge pull request #2460 from seefs001/feature/gemini-flash-minial

fix(gemini): handle minimal reasoning effort budget

relay/channel/gemini/adaptor.go

@@ -13,6 +13,7 @@ import (
 	relaycommon "github.com/QuantumNous/new-api/relay/common"
 	"github.com/QuantumNous/new-api/relay/constant"
 	"github.com/QuantumNous/new-api/setting/model_setting"
+	"github.com/QuantumNous/new-api/setting/reasoning"
 	"github.com/QuantumNous/new-api/types"
 
 	"github.com/gin-gonic/gin"
@@ -137,7 +138,7 @@ func (a *Adaptor) GetRequestURL(info *relaycommon.RelayInfo) (string, error) {
 			info.UpstreamModelName = strings.TrimSuffix(info.UpstreamModelName, "-thinking")
 		} else if strings.HasSuffix(info.UpstreamModelName, "-nothinking") {
 			info.UpstreamModelName = strings.TrimSuffix(info.UpstreamModelName, "-nothinking")
-		} else if baseModel, level := parseThinkingLevelSuffix(info.UpstreamModelName); level != "" {
+		} else if baseModel, level, ok := reasoning.TrimEffortSuffix(info.UpstreamModelName); ok && level != "" {
 			info.UpstreamModelName = baseModel
 		}
 	}

relay/channel/gemini/relay-gemini.go

@@ -98,6 +98,7 @@ func clampThinkingBudget(modelName string, budget int) int {
 // "effort": "high" - Allocates a large portion of tokens for reasoning (approximately 80% of max_tokens)
 // "effort": "medium" - Allocates a moderate portion of tokens (approximately 50% of max_tokens)
 // "effort": "low" - Allocates a smaller portion of tokens (approximately 20% of max_tokens)
+// "effort": "minimal" - Allocates a minimal portion of tokens (approximately 5% of max_tokens)
 func clampThinkingBudgetByEffort(modelName string, effort string) int {
 	isNew25Pro := isNew25ProModel(modelName)
 	is25FlashLite := is25FlashLiteModel(modelName)
@@ -118,18 +119,12 @@ func clampThinkingBudgetByEffort(modelName string, effort string) int {
 		maxBudget = maxBudget * 50 / 100
 	case "low":
 		maxBudget = maxBudget * 20 / 100
+	case "minimal":
+		maxBudget = maxBudget * 5 / 100
 	}
 	return clampThinkingBudget(modelName, maxBudget)
 }
 
-func parseThinkingLevelSuffix(modelName string) (string, string) {
-	base, level, ok := reasoning.TrimEffortSuffix(modelName)
-	if !ok {
-		return modelName, ""
-	}
-	return base, level
-}
-
 func ThinkingAdaptor(geminiRequest *dto.GeminiChatRequest, info *relaycommon.RelayInfo, oaiRequest ...dto.GeneralOpenAIRequest) {
 	if model_setting.GetGeminiSettings().ThinkingAdapterEnabled {
 		modelName := info.UpstreamModelName
@@ -186,7 +181,7 @@ func ThinkingAdaptor(geminiRequest *dto.GeminiChatRequest, info *relaycommon.Rel
 					ThinkingBudget: common.GetPointer(0),
 				}
 			}
-		} else if _, level := parseThinkingLevelSuffix(modelName); level != "" {
+		} else if _, level, ok := reasoning.TrimEffortSuffix(info.UpstreamModelName); ok && level != "" {
 			geminiRequest.GenerationConfig.ThinkingConfig = &dto.GeminiThinkingConfig{
 				IncludeThoughts: true,
 				ThinkingLevel:   level,

setting/reasoning/suffix.go

@@ -6,7 +6,7 @@ import (
 	"github.com/samber/lo"
 )
 
-var EffortSuffixes = []string{"-high", "-medium", "-low"}
+var EffortSuffixes = []string{"-high", "-medium", "-low", "-minimal"}
 
 // TrimEffortSuffix -> modelName level(low) exists
 func TrimEffortSuffix(modelName string) (string, string, bool) {