fix(kits): update Gemini models

kits/firestore-genai-chatbot/README.md

@@ -90,17 +90,17 @@ the CLI connects them to the function at deploy time.
 |---|---|---|---|---|
 | `provider` | `GENERATIVE_AI_PROVIDER` | no | `google-ai` | `google-ai` or `vertex-ai` |
 | `apiKey` | `API_KEY` | secret | — | Google AI API key |
-| `model` | `MODEL` | no | `gemini-2.5-flash` | Model id |
-| `vertexModelLocation` | `VERTEX_AI_MODEL_LOCATION` | no | `null` | Vertex model region |
+| `model` | `MODEL` | no | `gemini-3.6-flash` | Model id. Shape-checked at deploy time; not verified against the provider |
+| `vertexModelLocation` | `VERTEX_AI_MODEL_LOCATION` | no | `global` | Vertex model region. Gemini 3.x is only served on `global`, `us` and `eu` |
 | `collectionName` | `COLLECTION_NAME` | no | `generate` | Discussion collection |
 | `promptField` | `PROMPT_FIELD` | no | `prompt` | Prompt field name |
 | `responseField` | `RESPONSE_FIELD` | no | `response` | Response field name |
 | `orderField` | `ORDER_FIELD` | no | `createTime` | Ordering field |
 | `candidatesField` | `CANDIDATES_FIELD` | no | `candidates` | Candidates field name |
 | `context` | `CONTEXT` | no | (empty) | System context |
-| `temperature` | `TEMPERATURE` | no | (empty) | Sampling temperature |
-| `topP` | `TOP_P` | no | (empty) | Top-p |
-| `topK` | `TOP_K` | no | (empty) | Top-k |
+| `temperature` | `TEMPERATURE` | no | (empty) | Sampling temperature. Ignored by Gemini 3.x |
+| `topP` | `TOP_P` | no | (empty) | Top-p. Ignored by Gemini 3.x |
+| `topK` | `TOP_K` | no | (empty) | Top-k. Ignored by Gemini 3.x |
 | `candidateCount` | `CANDIDATE_COUNT` | no | `1` | Candidate count |
 | `maxOutputTokens` | `MAX_OUTPUT_TOKENS` | no | (empty) | Max output tokens |
 | `enableOverrides` | `ENABLE_DISCUSSION_OPTION_OVERRIDES` | no | `false` | Per-discussion option overrides |

kits/firestore-genai-chatbot/src/candidates.ts

@@ -0,0 +1,35 @@
+/*
+ * Copyright 2026 Google LLC
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    https://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/**
+ * Whether this configuration asks for multiple candidate responses.
+ *
+ * The client-selection gate and the Firestore write path must agree: the Genkit
+ * client only serves single-candidate configs, and writing the `candidates`
+ * field requires a field name to write it to. Keeping one predicate stops the
+ * two from drifting, which would either waste a multi-candidate request or send
+ * a single-candidate config down the legacy clients.
+ */
+export function wantsMultipleCandidates(config: {
+  candidateCount?: number;
+  candidatesField?: string;
+}): boolean {
+  return (
+    !!config.candidatesField &&
+    !!config.candidateCount &&
+    config.candidateCount > 1
+  );
+}

kits/firestore-genai-chatbot/src/config.ts

@@ -29,6 +29,13 @@ import {
   type SafetySetting,
 } from "./export-config";
 
+/**
+ * Shape check only — model ids are not validated against the provider, so an id
+ * that exists but is not served fails at request time. This catches typos like
+ * `gemini 3.6-flash` at deploy time instead of on every write.
+ */
+const MODEL_ID_PATTERN = /^[a-zA-Z0-9][a-zA-Z0-9.\-_/]*$/;
+
 const GENERATIVE_AI_PROVIDER_OPTIONS = ["google-ai", "vertex-ai"] as const;
 const VERTEX_MODEL_LOCATION_OPTIONS = [
   "null",
@@ -84,9 +91,28 @@ const params = {
     input: select([...GENERATIVE_AI_PROVIDER_OPTIONS]),
   }),
   apiKey: defineSecret("API_KEY"),
-  model: defineString("MODEL", { default: "gemini-2.5-flash" }),
+  model: defineString("MODEL", {
+    default: "gemini-3.6-flash",
+    input: {
+      text: {
+        example: "gemini-3.6-flash",
+        validationRegex: MODEL_ID_PATTERN.source,
+        validationErrorMessage:
+          "Model ids have no spaces, for example 'gemini-3.6-flash'.",
+      },
+    },
+  }),
+  /**
+   * Vertex AI location for the model. Defaults to `global` rather than the
+   * function region: Gemini 3.x is served on the `global`, `us` and `eu`
+   * endpoints only, so a single region such as `us-central1` returns 404 for the
+   * default `gemini-3.6-flash`. Set a specific region only with a model that is
+   * served there (for example a Gemini 2.5 model).
+   *
+   * @see https://cloud.google.com/vertex-ai/generative-ai/docs/learn/locations
+   */
   vertexModelLocation: defineString("VERTEX_AI_MODEL_LOCATION", {
-    default: "null",
+    default: "global",
     input: select([...VERTEX_MODEL_LOCATION_OPTIONS]),
   }),
   collectionName: defineString("COLLECTION_NAME", { default: "generate" }),
@@ -97,6 +123,12 @@ const params = {
     default: "candidates",
   }),
   context: defineString("CONTEXT", { default: "" }),
+  /**
+   * Sampling controls. Gemini 3.x deprecates `temperature`, `topP` and `topK`;
+   * the Vertex AI model card for `gemini-3.6-flash` states custom values are
+   * ignored. They still apply to Gemini 2.5 models, which retire in October
+   * 2026, so the params are kept for existing configurations.
+   */
   temperature: defineString("TEMPERATURE", { default: "" }),
   topP: defineString("TOP_P", { default: "" }),
   topK: defineString("TOP_K", { default: "" }),
@@ -129,6 +161,16 @@ const params = {
 /** The secret bound on the function so its value is available at runtime. */
 export const apiKeySecret = params.apiKey;
 
+/** Rejects a model id that cannot be a model id at all. */
+function requireModelId(model: string): string {
+  if (!MODEL_ID_PATTERN.test(model)) {
+    throw new Error(
+      `MODEL must be a model id with no spaces, for example 'gemini-3.6-flash'. Received: '${model}'`
+    );
+  }
+  return model;
+}
+
 /** Coerce an empty-string param value to `undefined`. */
 function optional(value: string): string | undefined {
   return value.length > 0 ? value : undefined;
@@ -164,7 +206,7 @@ export function configFromEnv(): GenaiChatbotConfig {
       (optional(params.provider.value()) as GenerativeAIProvider) ??
       GenerativeAIProvider.GOOGLE_AI,
     apiKey: params.apiKey.value(),
-    model: params.model.value(),
+    model: requireModelId(params.model.value()),
     vertexModelLocation:
       vertexModelLocation === "null" ? undefined : vertexModelLocation,
     projectId: getProjectId(),

kits/firestore-genai-chatbot/src/export-config.ts

@@ -45,7 +45,7 @@ export interface GenaiChatbotConfig {
   provider?: GenerativeAIProvider | "google-ai" | "vertex-ai";
   /** API key for the `google-ai` provider. */
   apiKey?: string;
-  /** Model id, e.g. `gemini-2.5-flash`. */
+  /** Model id, e.g. `gemini-3.6-flash`. */
   model: string;
   /** Vertex AI model location. */
   vertexModelLocation?: string;

kits/firestore-genai-chatbot/src/generate-chat-response.ts

@@ -15,6 +15,7 @@
  */
 
 import type { DocumentSnapshot } from "firebase-admin/firestore";
+import { wantsMultipleCandidates } from "./candidates";
 import type { ResolvedGenaiChatbotConfig } from "./export-config";
 import { fetchDiscussionOptions, fetchHistory } from "./firestore";
 import { getGenerativeClient } from "./generative-client";
@@ -53,12 +54,16 @@ export function createGenerateChatResponse(config: ResolvedGenaiChatbotConfig) {
       requestOptions = { ...requestOptions, ...discussionOptions };
     }
 
-    const shouldAddCandidatesField =
-      config.candidatesField &&
-      requestOptions.candidateCount &&
-      requestOptions.candidateCount > 1;
+    // Per-discussion overrides can raise the candidate count, so the client
+    // gate and this write decision must both use the effective value.
+    const candidateCount =
+      requestOptions.candidateCount ?? config.candidateCount;
+    const shouldAddCandidatesField = wantsMultipleCandidates({
+      candidateCount,
+      candidatesField: config.candidatesField,
+    });
 
-    const discussionClient = getGenerativeClient(config);
+    const discussionClient = getGenerativeClient(config, candidateCount);
     const result = await discussionClient.send(prompt, requestOptions);
     const response = result.response;
 

kits/firestore-genai-chatbot/src/generative-client/genkit.ts

@@ -30,6 +30,7 @@ import {
 } from "genkit";
 import { logger as genkitLogger } from "genkit/logging";
 import type { GenkitPluginV2 } from "genkit/plugin";
+import { wantsMultipleCandidates } from "../candidates";
 import type { ResolvedGenaiChatbotConfig } from "../export-config";
 import { logger } from "../logger";
 import {
@@ -103,58 +104,24 @@ export class GenkitDiscussionClient extends DiscussionClient<
     return genkit(genkitConfig);
   }
 
-  // TODO(migration): inherited verbatim from the legacy extension — this
-  // hardcoded model allowlist means new/custom/fine-tuned models need a package
-  // update. `googleAI.model()` / `vertexAI.model()` resolve any id dynamically;
-  // consider simplifying to that. Improvement, not a bug. Deferred from PR #431 review.
+  /**
+   * Resolves a Genkit model reference via `googleAI.model()` / `vertexAI.model()`.
+   * Any id is passed through so current Gemini releases work without a package update.
+   */
   static createModelReference(
     model: string,
     provider: string
   ): ModelReference<any> {
-    const modelReferences =
-      provider === "google-ai"
-        ? [
-            googleAI.model("gemini-1.5-flash"),
-            googleAI.model("gemini-1.5-pro"),
-            googleAI.model("gemini-2.0-flash"),
-            googleAI.model("gemini-2.0-flash-lite"),
-            googleAI.model("gemini-2.5-flash-lite"),
-            googleAI.model("gemini-2.5-flash"),
-            googleAI.model("gemini-2.5-pro"),
-            googleAI.model("gemini-3-pro-preview"),
-            googleAI.model("gemini-3-pro-image-preview"),
-          ]
-        : [
-            vertexAI.model("gemini-1.5-flash"),
-            vertexAI.model("gemini-1.5-pro"),
-            vertexAI.model("gemini-2.0-flash"),
-            vertexAI.model("gemini-2.0-flash-lite"),
-            vertexAI.model("gemini-2.0-flash-001"),
-            vertexAI.model("gemini-2.5-flash-lite"),
-            vertexAI.model("gemini-2.5-flash"),
-            vertexAI.model("gemini-2.5-pro"),
-            vertexAI.model("gemini-3-pro-preview"),
-            vertexAI.model("gemini-3-pro-image-preview"),
-          ];
-
-    const pluginName = provider === "google-ai" ? "googleai" : "vertexai";
-
-    for (const modelReference of modelReferences) {
-      if (modelReference.name === `${pluginName}/${model}`) {
-        return modelReference;
-      }
-      if (modelReference.info?.versions?.includes(model)) {
-        return modelReference.withVersion(model);
-      }
-    }
-    throw new Error("Model not found.");
+    return provider === "google-ai"
+      ? googleAI.model(model)
+      : vertexAI.model(model);
   }
 
   private createGenerateOptions(
     config: ResolvedGenaiChatbotConfig
   ): GenerateOptions {
     if (!config.model) {
-      throw new Error("Model not found.");
+      throw new Error("Model must be specified in the configuration.");
     }
 
     return {
@@ -172,17 +139,20 @@ export class GenkitDiscussionClient extends DiscussionClient<
     };
   }
 
-  /** Whether the Genkit client can serve this config (single candidate + known model). */
-  static shouldUseGenkitClient(config: ResolvedGenaiChatbotConfig): boolean {
-    const shouldReturnMultipleCandidates =
-      config.candidateCount && config.candidateCount > 1;
-    return (
-      !shouldReturnMultipleCandidates &&
-      !!GenkitDiscussionClient.createModelReference(
-        config.model,
-        config.provider
-      )
-    );
+  /**
+   * Whether the Genkit client can serve this request (single candidate).
+   *
+   * `candidateCount` is passed separately because per-discussion overrides can
+   * raise it above the deploy-time value.
+   */
+  static shouldUseGenkitClient(
+    config: ResolvedGenaiChatbotConfig,
+    candidateCount = config.candidateCount
+  ): boolean {
+    return !wantsMultipleCandidates({
+      candidateCount,
+      candidatesField: config.candidatesField,
+    });
   }
 
   async generateResponse(

kits/firestore-genai-chatbot/src/generative-client/google_ai.ts

@@ -17,6 +17,7 @@
 import { GoogleGenerativeAI, type SafetySetting } from "@google/generative-ai";
 import { logger } from "../logger";
 import { DiscussionClient, type Message } from "./base_class";
+import { answerText } from "./parts";
 
 interface GeminiChatOptions {
   history?: Message[];
@@ -106,7 +107,7 @@ export class GeminiDiscussionClient extends DiscussionClient<
         );
       });
 
-    const text = result.response.text();
+    const text = answerText(result.response.candidates?.[0]?.content?.parts);
 
     if (!text) {
       throw new Error("No text returned candidate");
@@ -115,8 +116,9 @@ export class GeminiDiscussionClient extends DiscussionClient<
     return {
       response: text,
       candidates:
-        result.response.candidates?.map((c) => c.content.parts[0].text ?? "") ??
-        [],
+        result.response.candidates?.map(
+          (c) => answerText(c.content.parts) ?? ""
+        ) ?? [],
       safetyMetadata: result.response.promptFeedback,
       history,
     };

kits/firestore-genai-chatbot/src/generative-client/index.ts

@@ -29,17 +29,20 @@ import { VertexDiscussionClient } from "./vertex_ai";
 type Client = Genkit | GoogleGenAI | GoogleGenerativeAI;
 
 /**
- * Selects and constructs the generative client for a resolved config. Prefers
- * the Genkit client when it can serve the request (single candidate, known
- * model), falling back to the provider-specific SDK clients.
+ * Selects and constructs the generative client for a request. Prefers the
+ * Genkit client when it can serve the request (single candidate), falling back
+ * to the provider-specific SDK clients.
  *
  * @param config - The resolved chatbot configuration.
+ * @param candidateCount - Effective candidate count for this request, which
+ *   per-discussion overrides can raise above the deploy-time value.
  * @returns A ready-to-use discussion client.
  */
 export const getGenerativeClient = (
-  config: ResolvedGenaiChatbotConfig
+  config: ResolvedGenaiChatbotConfig,
+  candidateCount = config.candidateCount
 ): DiscussionClient<Client, any, any> => {
-  if (GenkitDiscussionClient.shouldUseGenkitClient(config)) {
+  if (GenkitDiscussionClient.shouldUseGenkitClient(config, candidateCount)) {
     return new GenkitDiscussionClient(config);
   }
 

kits/firestore-genai-chatbot/src/generative-client/parts.ts

@@ -0,0 +1,34 @@
+/*
+ * Copyright 2026 Google LLC
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *    https://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+/** A response part as far as candidate parsing is concerned. */
+export interface TextPart {
+  text?: string;
+  thought?: boolean;
+}
+
+/**
+ * First non-thought text part of a candidate.
+ *
+ * Reading `parts[0].text` is not reliable for thinking models: they can lead
+ * with thought parts, or put the answer in a later part. Thought parts carry
+ * `thought: true`, which the pinned `@google/generative-ai` version does not
+ * type, so callers pass their own part shape in.
+ */
+export function answerText(parts?: TextPart[]): string | undefined {
+  return parts?.find((part) => !part.thought && typeof part.text === "string")
+    ?.text;
+}

kits/firestore-genai-chatbot/src/generative-client/vertex_ai.ts

@@ -22,6 +22,7 @@ import {
 } from "@google/genai";
 import { logger } from "../logger";
 import { DiscussionClient, type Message } from "./base_class";
+import { answerText } from "./parts";
 
 interface GeminiChatOptions {
   history?: Message[];
@@ -128,7 +129,7 @@ export class VertexDiscussionClient extends DiscussionClient<
     }
 
     const candidates = result.candidates
-      .map((candidate) => candidate.content?.parts?.[0]?.text)
+      .map((candidate) => answerText(candidate.content?.parts))
       .filter((text): text is string => typeof text === "string");
 
     if (candidates.length === 0) {