The case
A thousand reviews a week, and nobody reads them all. The store needs to know what the customer feels.
Isaac's store gets reviews in Portuguese and, now, orders from customers in Argentina. The team wants to separate complaints, find out what customers praise, and translate product descriptions without hiring one more person. None of these tasks requires a model trained from scratch.
This post builds on AI and Machine Learning, which explains the kinds of learning and the life cycle of a model. For the data that feeds these services, the Data and Analytics post shows where it lives.
The problem
Starting from scratch is expensive
Training your own model for each task requires labeled data, time and a team that knows how to evaluate the result. For most common tasks, Google Cloud already has a ready model that solves the problem.
Three paths
From the most ready to the most custom
Google Cloud offers three paths to use AI, from the one that takes the least work to the one that gives the most control.
| Path | Who uses it | What you do |
|---|---|---|
| Pre-trained APIs | Those who want a common task solved | Call the service over the internet, without training anything |
| BigQuery ML | Those who already have the data in BigQuery and know SQL | Train and use models with SQL commands, where the data is |
| Vertex AI | Those who need a custom model or generative AI with control | Train, evaluate, publish and monitor models, with access to Gemini models |
Pre-trained APIs
One common task, one call
Pre-trained APIs cover tasks like understanding the meaning of a text, translating, transcribing audio and reading documents. You send the content and get the answer, without worrying about the model behind it.
Text analysis (Natural Language)
Finds out whether a review is positive or negative, and which entities appear in it, such as brands and products.
Translation
Translates a product description into Spanish before opening the store in Argentina.
Document reading (Document AI)
Reads invoices and extracts fields such as the total amount, the date and the supplier, without typing anything.
BigQuery ML
A model with SQL, where the data is
BigQuery ML lets you train, evaluate and use models with SQL commands, directly in the data warehouse. Whoever already queries BigQuery does not need to move the data anywhere else.
At the store, the sales team can forecast next week's demand by category using the sales that are already in BigQuery. The result comes back as a table, ready for the report.
Vertex AI
The platform for custom models
Vertex AI brings together, in one place, what you need to train, evaluate, publish and monitor models. It gives access to Gemini models and to a catalog of ready models, Model Garden, with models from Google, open models and third-party models.
Custom training
You bring your data and train a model for your problem, with control over each step.
Generative AI
Gemini models generate text, summarize documents and answer questions. You can tune them (fine-tuning) with your data when the standard answer is not enough.
Monitoring
Checks whether the model keeps getting it right, and can warn when the data changes its pattern, if monitoring is set up, as in the life cycle of a model.
In practice
Choosing the path
Start with the simplest option that solves the problem. Move up to the next level only when the previous one is not enough.
Common task
Use a pre-trained API. If it solves the problem, there is nothing to train.
Forecast with data in BigQuery
Use BigQuery ML, if the team knows SQL and the model is simple.
Tailored model
Use Vertex AI, when the problem requires a custom model or a tuned generative AI.
Key terms
Five terms to remember
Tap a card to see the definition.
Quick test
Three questions, exam style
01
The store wants to translate the product descriptions into Spanish, without training any model. Which path should it use?
02
The analysts want to forecast demand using the sales already in BigQuery, writing SQL queries. Which path should they use?
03
The store needs a tailored model, with monitoring and a generative assistant tuned to its data. Which path should it use?
In any cloud exam, the question is usually which path solves the problem with the least work. The order is always the same: first the pre-trained API, then SQL where the data already is, and only then a custom model.