Gemini - Create column description

Overview

This guide walks you through setting up the Gemini AI Column Description Generator in a new Google Cloud Platform (GCP) project. This tool automatically generates descriptions for database columns using Google's Gemini AI model integrated with BigQuery ML.

Step 1: Enable Required APIs

The following APIs need to be activated

BigQuery API - For data processing and ML operations
Vertex AI API - For Gemini AI model access gcloud services enable aiplatform.googleapis.com
Cloud Resource Manager API - For project management

Step 2: Set Up IAM Roles and Permissions

Required Roles for Script Execution

The service account or user running the script needs the following roles:

BigQuery Roles

BigQuery Data Editor (roles/bigquery.dataEditor)
- Create and modify datasets and tables
- Execute queries and manage temporary tables
BigQuery Job User (roles/bigquery.jobUser)
- Run BigQuery jobs and queries
BigQuery Connection User (roles/bigquery.connectionUser)
- Use BigQuery connections for ML models

Vertex AI Roles

Vertex AI User (roles/aiplatform.user)
- Access Vertex AI services and models
ML Developer (roles/ml.developer)
- Use ML models in BigQuery ML

Additional Roles

Service Usage Consumer (roles/serviceusage.serviceUsageConsumer)
- Use enabled Google Cloud services

Assign Roles via Console

Go to IAM & Admin > IAM
Find your user account or service account
Click "Edit principal"
Add each required role listed above

Assign Roles via gcloud CLI

# Replace [USER_EMAIL] with the actual email
export USER_EMAIL="your-email@domain.com"
export PROJECT_ID="your-project-id"

gcloud projects add-iam-policy-binding $PROJECT_ID \
    --member="user:$USER_EMAIL" \
    --role="roles/bigquery.dataEditor"

gcloud projects add-iam-policy-binding $PROJECT_ID \
    --member="user:$USER_EMAIL" \
    --role="roles/bigquery.jobUser"

gcloud projects add-iam-policy-binding $PROJECT_ID \
    --member="user:$USER_EMAIL" \
    --role="roles/bigquery.connectionUser"

gcloud projects add-iam-policy-binding $PROJECT_ID \
    --member="user:$USER_EMAIL" \
    --role="roles/aiplatform.user"

gcloud projects add-iam-policy-binding $PROJECT_ID \
    --member="user:$USER_EMAIL" \
    --role="roles/ml.developer"

Step 3: Create BigQuery Datasets

Create Source Dataset

Create a dataset to hold your source tables (if not already existing):

CREATE SCHEMA IF NOT EXISTS `your-project-id.DM`
OPTIONS (
  description = "Data Mart - Source tables for analysis",
  location = "US"  -- Choose appropriate location
);

Create Output Dataset

Create a dataset for storing the AI-generated descriptions:

CREATE SCHEMA IF NOT EXISTS `your-project-id.TEST_GEMINI`
OPTIONS (
  description = "Gemini AI generated column descriptions",
  location = "US"  -- Must match source dataset location
);

Step 4: Set Up Vertex AI Connection

Create Vertex AI Connection in BigQuery (documentation)

Via BigQuery Console:
- Go to BigQuery in Google Cloud Console
- Click + Add data, and then use the search bar for data sources to search for Vertex AI. Click on the result for Vertex AI then click Vertex AI Models: BigQuery Federation.
- Select Connection type as Vertex AI remote models, remote functions and BigLake (Cloud Resource) if not already selected and set Connection ID to vertex-ai-connection.

Get Connection Service Account

After creating the connection, note the service account email that's automatically created. You'll need to grant it permissions.

-- Get connection details
SELECT * FROM `your-project-id.us-central1.INFORMATION_SCHEMA.CONNECTIONS`
WHERE connection_name = 'vertex-ai-connection';

Grant Vertex AI Access to Connection Service Account

The connection service account needs access to Vertex AI:

# Get the service account email from the connection details above
export CONNECTION_SA="service-account-email@gcp-sa-bigquery-condel.iam.gserviceaccount.com"

gcloud projects add-iam-policy-binding $PROJECT_ID \
    --member="serviceAccount:$CONNECTION_SA" \
    --role="roles/aiplatform.user"

Step 5: Create Gemini AI Model in BigQuery

Create Remote Model

CREATE OR REPLACE MODEL `your-project-id.YOUR_DATASET.gemini_remote_model`
REMOTE WITH CONNECTION `your-project-id.europe-west1.vertex-ai-connection`
OPTIONS (
  ENDPOINT = 'gemini-2.0-flash'
);

Test Model Connection

-- Test the model with a simple query
SELECT ml_generate_text_llm_result
FROM ML.GENERATE_TEXT(
  MODEL `your-project-id.TEST_GEMINI.gemini_remote_model`,
  (SELECT "Test connection to Gemini AI" AS prompt),
  STRUCT(TRUE AS flatten_json_output)
);

Step 6: Configure and Deploy the Script

Update Configuration Variables

In the main script, update these variables with your project details:

-- Update these variables in the script
DECLARE dataset_name STRING DEFAULT 'your-project-id.DM';
DECLARE output_dataset STRING DEFAULT 'TEST_GEMINI';
DECLARE model_name STRING DEFAULT 'TEST_GEMINI.gemini_remote_model';
DECLARE sample_size INT64 DEFAULT 5;

Upload Sample Data (Optional)

If you don't have existing tables, create some sample tables for testing:

-- Create a sample table for testing
CREATE OR REPLACE TABLE `your-project-id.DM.sample_customers` AS
SELECT 
  ROW_NUMBER() OVER() as customer_id,
  CONCAT("Customer_", CAST(ROW_NUMBER() OVER() AS STRING)) as customer_name,
  ["New York", "Los Angeles", "Chicago", "Houston", "Phoenix"][SAFE_ORDINAL(MOD(ROW_NUMBER() OVER(), 5) + 1)] as city,
  RAND() * 10000 as purchase_amount,
  DATE_ADD("2024-01-01", INTERVAL CAST(RAND() * 365 AS INT64) DAY) as purchase_date
FROM UNNEST(GENERATE_ARRAY(1, 100)) as n;

Step 7: Execute the Script

Open BigQuery Console
Navigate to your project
Paste the complete script in the query editor
Update the configuration variables at the top
Run the script

Step 8: Verify Installation

Check Results

-- Verify the output table was created
SELECT 
  table_name,
  COUNT(*) as column_count
FROM `your-project-id.TEST_GEMINI.column_descriptions_with_ai`
GROUP BY table_name;

-- View sample results
SELECT * FROM `your-project-id.TEST_GEMINI.column_descriptions_with_ai` LIMIT 5;

Troubleshooting

Common Issues

"Permission denied" errors
- Verify all required APIs are enabled
- Check IAM roles are correctly assigned
- Ensure connection service account has Vertex AI access
"Model not found" errors
- Verify the Vertex AI connection was created successfully
- Check the model name matches exactly
- Ensure the connection region matches your dataset location
"No tables found" errors
- Verify source dataset exists and contains tables
- Check dataset names in the configuration variables
- Ensure the service account has access to source datasets
Quota exceeded errors
- Check Vertex AI quotas in the Cloud Console
- Consider reducing sample_size or processing fewer tables at once

Monitoring and Costs

Monitor usage: Go to Cloud Console > Vertex AI > Quotas to track API usage
Cost estimation: Gemini AI pricing varies by model and token usage
BigQuery costs: Query processing and storage costs apply

Security Best Practices

Use least privilege: Grant only the minimum required permissions
Service accounts: Consider using dedicated service accounts for production
Data access: Implement row-level security for sensitive datasets
Audit logging: Enable Cloud Audit Logs for compliance
Network security: Use VPC Service Controls for additional isolation

Next Steps

Once installed, you can:

Schedule the script to run periodically using Cloud Scheduler
Integrate with data catalog tools for metadata management
Customize prompts for domain-specific descriptions
Extend to other Vertex AI models for different use cases

Support

For issues specific to:

Last updated: June 2025

Page tree