feat(js): thinking mode options for Gemini models (#2851)

Author: EduardWayland

js/plugins/googleai/src/gemini.ts

@@ -17,8 +17,8 @@
 import {
   EnhancedGenerateContentResponse,
   FileDataPart,
-  FunctionCallingMode,
   FunctionCallPart,
+  FunctionCallingMode,
   FunctionDeclaration,
   FunctionResponsePart,
   GenerateContentCandidate as GeminiCandidate,
@@ -37,27 +37,27 @@ import {
   ToolConfig,
 } from '@google/generative-ai';
 import {
-  Genkit,
   GENKIT_CLIENT_HEADER,
+  Genkit,
   GenkitError,
   JSONSchema,
   z,
 } from 'genkit';
 import {
   CandidateData,
   GenerationCommonConfigSchema,
-  getBasicUsageStats,
   MediaPart,
   MessageData,
   ModelAction,
   ModelInfo,
   ModelMiddleware,
-  modelRef,
   ModelReference,
   Part,
   ToolDefinitionSchema,
   ToolRequestPart,
   ToolResponsePart,
+  getBasicUsageStats,
+  modelRef,
 } from 'genkit/model';
 import {
   downloadRequestMedia,
@@ -132,6 +132,29 @@ export const GeminiConfigSchema = GenerationCommonConfigSchema.extend({
         "'gemini-2.0-flash-exp' model at present."
     )
     .optional(),
+  thinkingConfig: z
+    .object({
+      includeThoughts: z
+        .boolean()
+        .describe(
+          'Indicates whether to include thoughts in the response.' +
+            'If true, thoughts are returned only when available.'
+        )
+        .optional(),
+      thinkingBudget: z
+        .number()
+        .min(0)
+        .max(24576)
+        .describe(
+          'The thinking budget parameter gives the model guidance on the ' +
+            'number of thinking tokens it can use when generating a response. ' +
+            'A greater number of tokens is typically associated with more detailed ' +
+            'thinking, which is needed for solving more complex tasks. ' +
+            'Setting the thinking budget to 0 disables thinking.'
+        )
+        .optional(),
+    })
+    .optional(),
 }).passthrough();
 export type GeminiConfig = z.infer<typeof GeminiConfigSchema>;
 

js/plugins/vertexai/src/gemini.ts

@@ -267,6 +267,29 @@ export const GeminiConfigSchema = GenerationCommonConfigSchema.extend({
         'With NONE, the model is prohibited from making function calls.'
     )
     .optional(),
+  thinkingConfig: z
+    .object({
+      includeThoughts: z
+        .boolean()
+        .describe(
+          'Indicates whether to include thoughts in the response.' +
+            'If true, thoughts are returned only when available.'
+        )
+        .optional(),
+      thinkingBudget: z
+        .number()
+        .min(0)
+        .max(24576)
+        .describe(
+          'The thinking budget parameter gives the model guidance on the ' +
+            'number of thinking tokens it can use when generating a response. ' +
+            'A greater number of tokens is typically associated with more detailed ' +
+            'thinking, which is needed for solving more complex tasks. ' +
+            'Setting the thinking budget to 0 disables thinking.'
+        )
+        .optional(),
+    })
+    .optional(),
 }).passthrough();
 
 /**