Limit thinking budget to 128 (smallest possible for gemini pro)

Author: jahooma

backend/src/llm-apis/vercel-ai-sdk/ai-sdk.ts

@@ -63,6 +63,7 @@ export const promptAiSdkStream = async function* (
     model: Model
     userId: string | undefined
     chargeUser?: boolean
+    thinkingBudget?: number
   } & Omit<Parameters<typeof streamText>[0], 'model'>
 ) {
   const startTime = Date.now()
@@ -75,6 +76,7 @@ export const promptAiSdkStream = async function* (
       google: {
         thinkingConfig: {
           includeThoughts: true,
+          thinkingBudget: options.thinkingBudget ?? 128,
         },
       } satisfies GoogleGenerativeAIProviderOptions,
     },

backend/src/prompt-agent-stream.ts

@@ -17,11 +17,13 @@ export const getAgentStream = (params: {
   fingerprintId: string
   userInputId: string
   userId: string | undefined
+  thinkingBudget?: number
 }) => {
   const {
     costMode,
     selectedModel,
     stopSequences,
+    thinkingBudget,
     clientSessionId,
     fingerprintId,
     userInputId,
@@ -65,7 +67,7 @@ export const getAgentStream = (params: {
       }
       if (!options.providerOptions.gemini.thinkingConfig) {
         options.providerOptions.gemini.thinkingConfig = {
-          thinkingBudget: 0,
+          thinkingBudget: thinkingBudget ?? 128,
         }
       }
     }