Gemini Enterprise Demo Environment

This module prepares a Google Cloud project for an instructor-led demonstration of Gemini Enterprise, built around the fictional Cymbal Pools company. It is a standalone module: it creates its own bucket, dataset, agent and app, and does not depend on any shared foundation infrastructure.

On apply the module creates the Gemini Enterprise app with Google Identity configured, a searchable data store of Cymbal Pools documents, a content bucket (Drive documents, announcement image, pH-table image, BigQuery seed data, ADK agent source), a BigQuery installation_requests table, a custom ADK BigQuery agent on Vertex AI Agent Runtime with the IAM it needs, and a Model Armor template. The instructor then completes the handful of steps that need an interactive Google Workspace sign-in (Drive/Calendar connectors, OAuth client, agent registration) and runs the demos.

This guide covers the cloud services the module provisions and how to explore and operate them from the Google Cloud Console and the command line. The full pre-class checklist and in-class demo script are in the Lab Guide.

Resources carry the deployment suffix <id>: bucket <project>-ge-<id>, app cymbal-pools-ge-<id>, data store cymbal-pools-docs-<id>, Model Armor template cymbal-pools-ma-<id>, and Agent Runtime engine "BigQuery Pool Data Agent (<id>)".

---

1. Overview

Capability	Google Cloud service	Notes
Agentic workspace app	Gemini Enterprise (Discovery Engine)	APP_TYPE_INTRANET app, Enterprise search tier with the LLM add-on, company name "Cymbal Pools"
Identity	Gemini Enterprise ACL config	Google Identity (GSUITE) selected for the app's location
Company documents	Discovery Engine data store	Cloud Storage-backed, unstructured; indexes the brochure PDF and analysis DOCX
Demo content	Cloud Storage	Drive documents, announcement image, pH-table image, seed CSV, adk_to_ge/ agent source
Pool installation data	BigQuery	cymbal_pools.installation_requests, seeded with 60 sample rows
Custom agent	Vertex AI Agent Runtime (Agent Engine) + ADK	BigQuery agent deployed with adk deploy agent_engine, running as resource_creator_identity
AI safety	Model Armor	Template with SDP, RAI, prompt-injection/jailbreak and malicious-URI filters

Things to know up front:

• Gemini Enterprise must be activated in the project first. On a project that has never used it, open Gemini Enterprise in the console once and click Start free trial. Enabling discoveryengine.googleapis.com is not the same thing, and app creation fails until the trial is started. Set create_gemini_enterprise_app = false to create the app by hand instead.
• Several steps are deliberately manual. The Google Drive and Calendar connectors and the OAuth consent screen/client need interactive OAuth flows that only the console provides. Registering the BigQuery agent needs that OAuth client, and granting the agent's "Agent User" role to all users has no API. Announcements, Feature Management toggles and attaching Model Armor to the assistant are demoed live.
• Everything shares one Gemini Enterprise location. The app, its data stores, the identity provider, the Drive/Calendar connectors and the agent registration must all live in ge_location (global by default).
• The agent deploy runs Python on the Terraform runner. null_resource.deploy_adk_agent creates a virtualenv under $HOME/.local/ge-adk-venv, installs google-adk, and runs adk deploy agent_engine through a small wrapper (scripts/adk_deploy.py) that makes the ADK impersonate resource_creator_identity. The deploy takes 5–10 minutes.
• The demo content is original. Cymbal Pools is fictional; the PDF, DOCX, images and CSV in assets/ were written for this module and can be regenerated with assets/generate_assets.py.

---

2. Google Cloud Services & How to Explore Them

All commands assume these are set to match your deployment. Resource names are reported in the deployment Outputs.

export PROJECT="<project_id>"
export REGION="us-central1"      # region (Agent Runtime)
export GE_LOCATION="global"      # ge_location
export ID="<deployment_id>"
export GE_HOST=discoveryengine.googleapis.com   # use ${GE_LOCATION}-discoveryengine.googleapis.com for us/eu

A. Gemini Enterprise app, identity and data store

The app (cymbal-pools-ge-<id>) is an intranet-type engine with the Enterprise search tier and LLM add-on. Google Identity is set as the identity provider for ge_location so ACL-enforced connectors (Drive, Calendar) can be created there. The cymbal-pools-docs-<id> data store is attached to the app at creation and indexes the documents under gs://<bucket>/drive/.

• Console: Gemini Enterprise → Apps (set the location filter to ge_location or All locations) → Cymbal Pools GE. Connected data stores, Integration, Agents and Configurations are in the app's left nav.
• CLI:

  TOKEN=$(gcloud auth print-access-token)
  # The app
  curl -s -H "Authorization: Bearer $TOKEN" -H "X-Goog-User-Project: $PROJECT" \
    "https://$GE_HOST/v1/projects/$PROJECT/locations/$GE_LOCATION/collections/default_collection/engines/cymbal-pools-ge-$ID"
  # Data stores in the location (the Drive/Calendar/Announcements ones appear here once created)
  curl -s -H "Authorization: Bearer $TOKEN" -H "X-Goog-User-Project: $PROJECT" \
    "https://$GE_HOST/v1/projects/$PROJECT/locations/$GE_LOCATION/collections/default_collection/dataStores"
  # Documents indexed from the bucket
  curl -s -H "Authorization: Bearer $TOKEN" -H "X-Goog-User-Project: $PROJECT" \
    "https://$GE_HOST/v1/projects/$PROJECT/locations/$GE_LOCATION/collections/default_collection/dataStores/cymbal-pools-docs-$ID/branches/default_branch/documents"

B. Cloud Storage — demo content

Object	Used for
drive/Cymbal Pools Convention Brochure.pdf	Upload to Google Drive; "Summarize @" file-targeting demo
drive/Cymbal Pools Pool Installation Analysis.docx	Upload to Google Drive; "how do we conduct a pool site inspection?" enterprise search demo
pool party.png	Image URL of the "Annual pool party" announcement (public-read when public_announcement_image = true)
pH table.png	Paste into the omnibar for the "Convert this image to tabular data" demo
bigquery/installation_requests.csv	Seed data loaded into BigQuery
adk_to_ge/	ADK agent source (bigquery_agent/, construct_auth_uri.py, requirements.txt)

• Console: Cloud Storage → Buckets → <project>-ge-<id> (direct link in the drive_documents_console_url output).
• CLI:

  gcloud storage ls -r "gs://$PROJECT-ge-$ID/"
  gcloud storage cp "gs://$PROJECT-ge-$ID/drive/*" .           # the two Drive documents
  gcloud storage cp -r "gs://$PROJECT-ge-$ID/adk_to_ge" .      # the agent source
  curl -sI "https://storage.googleapis.com/$PROJECT-ge-$ID/pool%20party.png" | head -1   # expect 200 if public

C. BigQuery — installation requests

cymbal_pools.installation_requests holds residential pool installation requests: request_id, request_date, length_m, width_m, depth_m, has_hot_tub, has_waterfall, zip_code, customer_phone, customer_email. It is seeded by a BigQuery load job whose ID embeds the CSV's hash, so rows the agent inserts during a demo survive later applies unless the seed CSV itself changes.

• Console: BigQuery → Explorer → <project> → cymbal_pools → installation_requests. Queries the agent runs appear under Job history → Project history.
• CLI:

  bq show --schema --format=prettyjson "$PROJECT:cymbal_pools.installation_requests"
  bq query --use_legacy_sql=false \
    "SELECT COUNT(*) AS requests, ROUND(AVG(length_m*width_m*depth_m),1) AS avg_volume_m3 FROM \`$PROJECT.cymbal_pools.installation_requests\`"

D. Vertex AI Agent Runtime — BigQuery agent

The ADK agent (adk_to_ge/bigquery_agent) uses the ADK BigQueryToolset with write mode allowed and instructions that restrict it to SELECT/INSERT on installation_requests. By default it queries as the AI Platform Reasoning Engine service agent (service-<number>@gcp-sa-aiplatform-re.iam.gserviceaccount.com), which the module grants roles/aiplatform.user, roles/bigquery.user and roles/bigquery.dataEditor. Setting agent_auth_id makes it use the Gemini Enterprise user's OAuth token instead (see Behaviour).

• Console: Vertex AI → Agent Runtime (select region) → "BigQuery Pool Data Agent (<id>)".
• CLI:

  TOKEN=$(gcloud auth print-access-token)
  curl -s -H "Authorization: Bearer $TOKEN" \
    "https://$REGION-aiplatform.googleapis.com/v1/projects/$PROJECT/locations/$REGION/reasoningEngines" \
    | python3 -c "import json,sys; [print(e['name'], '|', e.get('displayName')) for e in json.load(sys.stdin).get('reasoningEngines',[])]"
  gcloud projects get-iam-policy "$PROJECT" --flatten="bindings[].members" \
    --filter="bindings.members:gcp-sa-aiplatform-re" --format="table(bindings.role)"

E. Model Armor

One template, cymbal-pools-ma-<id>, in model_armor_location (defaults to ge_location, except that a global app gets a us template: Model Armor rejects templates in global with UNSUPPORTED_REQUEST_LOCATION, and a global app's assistant accepts a us template, both confirmed live). Each filter maps to a group of the lab's test prompts: basic Sensitive Data Protection (card numbers, credentials), responsible-AI filters (harassment, hate speech, dangerous, sexually explicit) at model_armor_confidence, prompt-injection/jailbreak detection, and malicious-URI detection. Enforcement is INSPECT_AND_BLOCK. The module does not attach it to the assistant; the instructor does that live under Configurations → Assistant → Enable Model Armor.

• Console: Security → Model Armor → cymbal-pools-ma-<id>.
• CLI:

  gcloud model-armor templates describe "cymbal-pools-ma-$ID" --location="$GE_LOCATION" --project="$PROJECT"

---

3. Behaviour

What gets provisioned on apply. In dependency order: the project APIs (Discovery Engine, Vertex AI, BigQuery, Cloud Storage, IAM and IAM Credentials, Resource Manager, Service Usage, Model Armor, Cloud Build, Logging, Monitoring, Cloud Trace); the content bucket and its objects; the BigQuery dataset, table and seed load job; the Google Identity ACL config; the Discovery Engine service agent (with read access to the bucket), the document data store, and a documents:import call; the Gemini Enterprise app; the Model Armor template; and finally the agent deploy followed by the Reasoning Engine service agent's IAM bindings. Those bindings must come after the deploy because that service agent only exists once the project's first engine has been created.

Document import. Discovery Engine has no Terraform resource for importing documents, so null_resource.import_cymbal_docs calls documents:import with gs://<bucket>/drive/* and retries for up to about two minutes while the service agent's bucket grant propagates. The import is asynchronous: apply returns once it has started, and indexing finishes a few minutes later. The import re-runs when either document's content changes.

Agent deploy and identity. adk deploy agent_engine has no impersonation option, so the module runs it through scripts/adk_deploy.py. That wrapper replaces google.auth.default() with short-lived impersonated credentials for resource_creator_identity before the ADK loads, so the engine is created by the same identity as every other resource. With resource_creator_identity empty it behaves exactly like adk. The wrapper writes the resulting reasoningEngines resource name to scripts/reasoning_engine.txt, and the reasoning_engine output reads that file.

Model location vs. deploy region. adk deploy --region X forces GOOGLE_CLOUD_LOCATION=X inside the deployed container, overriding any value in the agent's .env. (The original lab's GOOGLE_CLOUD_LOCATION=global line is therefore ignored.) The module passes the model location separately as MODEL_LOCATION (agent_model_location, default global), and the agent applies it before the model client is created.

Per-user vs. service-agent BigQuery access. With agent_auth_id empty (default) the agent queries as the Reasoning Engine service agent, so the demo works whether or not the agent was registered with an OAuth authorization. Set agent_auth_id to the Gemini Enterprise Authorization ID (e.g. bq-auth) to have queries run as the signed-in user; the agent then reads the token Gemini Enterprise stores in session state under that key. Changing it redeploys the agent.

Console changes are preserved. The app ignores drift on data_store_ids, features and knowledge_graph_config. As a result, connectors the instructor attaches and Feature Management toggles (Agent Designer, image model, session sharing) are not reverted by a later apply.

Manual follow-up the instructor performs. Upload the Drive documents, create the Drive/Calendar connectors and the Announcements data store, create the OAuth consent screen and Web client, register the BigQuery agent (optionally with an Authorization), grant All Users the Agent User role, confirm Workforce Identity on the app's Integration page, and enable Feature Management options. All of these are listed step by step in the Lab Guide.

Cleanup behaviour. Destroy removes the app, data store, ACL config, bucket (force-destroyed with its contents), BigQuery dataset (with contents), Model Armor template, IAM bindings, and the Agent Runtime engine. The engine's destroy provisioner finds it by display name rather than the local file, since that file does not survive between platform runs. Destroy does not remove anything created by hand. Deleting the app does not delete the data stores attached to it, so the Drive, Calendar and Announcements data stores must be deleted from Gemini Enterprise → Data stores afterwards. The same applies to the OAuth client and consent screen, the Authorization resource, and the files uploaded to Google Drive. The agent registration lives inside the app and goes with it. APIs are left enabled.

---

4. Configuration Variables

Variables are grouped exactly as they appear on the deployment platform.

Group 1 — Project & Location

Variable	Default	Description
project_id	(required)	Target project. Must already exist. For Qwiklabs, the lab project with the student account, so connectors and the app share one Workspace identity.
region	us-central1	Agent Runtime region for the ADK agent. Must be permitted by any constraints/gcp.resourceLocations policy.

Group 2 — Gemini Enterprise App

Variable	Default	Description
create_gemini_enterprise_app	true	Create the app and the Cymbal Pools document data store. Needs the free trial (or a license) started first.
ge_location	global	global, us or eu. App, data stores, IdP, connectors and agent registration must all share it.
app_display_name	Cymbal Pools GE	App display name.
company_name	Cymbal Pools	Company name on the app (Advanced Options).
configure_google_identity	true	Select Google Identity as the IdP for ge_location. Required before Drive/Calendar connectors.

Group 3 — Demo Content

Variable	Default	Description
bucket_location	US	Location of the content bucket.
public_announcement_image	true	Make only pool party.png public-read so the announcement renders for every user.
bq_dataset_id	cymbal_pools	BigQuery dataset for installation_requests.
bq_location	US	BigQuery dataset location.

Group 4 — Custom ADK Agent

Variable	Default	Description
deploy_adk_agent	true	Deploy the BigQuery agent to Agent Runtime and grant its service agent Vertex AI User, BigQuery User and BigQuery Data Editor.
agent_display_name	BigQuery Pool Data Agent	Agent Runtime display name; the deployment ID is appended.
agent_model	gemini-3.5-flash	Model the agent uses. Changing it redeploys the agent as a new engine.
agent_model_location	global	Vertex AI location for model calls.
agent_auth_id	(empty)	Gemini Enterprise Authorization ID; when set, queries run as the signed-in user.

Group 5 — Model Armor

Variable	Default	Description
create_model_armor_template	true	Create the demo Model Armor template.
model_armor_location	(empty → ge_location, with global → us)	Template location. Model Armor has no global templates.
model_armor_confidence	MEDIUM_AND_ABOVE	Blocking threshold for RAI and prompt-injection filters (LOW_AND_ABOVE, MEDIUM_AND_ABOVE, HIGH).

---

5. Outputs

Output	Description
deployment_id	The deployment suffix used in resource names.
project_id	GCP project ID.
demo_bucket	Content bucket name.
drive_documents_console_url	Console page listing the two documents to upload to Google Drive.
agent_source_copy_cmd	gcloud storage cp -r command to fetch adk_to_ge/ into Cloud Shell.
announcement_image_url	Image URL for the "Annual pool party" announcement.
ph_table_image_url	Authenticated URL of the pH table image for the image-to-table demo.
gemini_enterprise_location	The location to use for connectors and agent registration.
gemini_enterprise_app_id	Engine ID of the app (not created when disabled).
gemini_enterprise_console_url	Gemini Enterprise console page for the project.
cymbal_docs_data_store	Resource name of the Cymbal Pools Documents data store.
bigquery_table	Fully qualified installation_requests table.
reasoning_engine	Agent Runtime resource name to paste into Agents → Add agent.
model_armor_template	Template resource name to paste into Configurations → Assistant → Enable Model Armor.
oauth_redirect_uris	The two redirect URIs to add to the Web OAuth client used by the connectors and the agent's authorization.

---

6. Configuration Pitfalls & Sensible Defaults

Risk: Critical (data loss / outage / security) — High (service degraded) — Medium (cost or partial degradation) — Low (minor).

Setting	Sensible value	Risk	Consequence if wrong
deployment_id	set once	Critical	Embedded in every resource name. Changing it recreates the app (losing hand-attached connectors and agent registrations), bucket, data store and agent.
create_gemini_enterprise_app	true	High	If app creation fails with a licensing error, Gemini Enterprise is not activated in the project: start the free trial in the console and redeploy, or set false and create the app by hand. (Not needed on the Qwiklabs project this was tested against.)
agent_model	a model allowed by constraints/vertexai.allowedModels	High	Lab projects can restrict or deny all Vertex AI models. The agent still deploys, but every query fails with FAILED_PRECONDITION ... disallowed Gen AI model. Check the effective policy before class.
ge_location	global (or us on quota-limited sandboxes)	High	Connectors and the agent registration must be in the same location as the app, and a mismatch is rejected. Some sandbox projects have a custom-agent quota of 0 at global; use us there. Changing it recreates the app.
configure_google_identity	true	High	Without an IdP for the location, the Drive/Calendar connector wizard fails with "IdP must be selected before creating an ACLed Data Connector".
region	a region allowed by org policy	High	constraints/gcp.resourceLocations can reject Agent Runtime in the default region with a 412; pick an allowed region.
public_announcement_image	true, or false under public-access prevention	Medium	With constraints/storage.publicAccessPrevention enforced, the public ACL fails with a 412. Set false and use an image hosted elsewhere for the announcement.
agent_model / agent_auth_id	set before registering the agent	Medium	Either change redeploys the agent as a new engine, so the Gemini Enterprise registration pointing at the old reasoning_engine must be re-created.
agent_auth_id	empty unless demoing per-user access	Medium	When set, queries run as the signed-in user, who then needs BigQuery access on the dataset; an unregistered or mismatched Authorization ID gives the agent no credentials.
enable_services	true	High	If the required APIs are not already enabled and this is false, resource creation fails immediately.
model_armor_confidence	MEDIUM_AND_ABOVE	Low	HIGH may let some of the lab's milder test prompts through; LOW_AND_ABOVE may block ordinary demo prompts.

---

For the full instructor walkthrough (pre-class checklist, connector and agent registration steps, and the in-class demo prompts), see the Lab Guide.