PMLE - Architecting Low-Code AI Solutions (12% of the exam) - Section 1.2

Build AI solutions with Google Cloud AI APIs and foundation models, selecting models from Model Garden, using industry-specific APIs such as Document AI, Vision, and Translate, tuning Gemini, Imagen, and Veo, and optimising for cost, latency, and availability.

Select and tune foundation models from Model Garden, and apply industry-specific APIs such as Document AI, Vision, and Translate to avoid building custom models from scratch. Weigh cost, latency, and availability trade-offs when choosing between calling a Gemini or Imagen endpoint and hosting a self-managed model.

Model GardenDocument AI APIGemini and ImagenCost and latency optimisation

Practice question for this objective

Free sampleArchitecting Low-Code AI Solutionsmedium

A global retailer is building a pipeline to process photographs that shoppers upload with product reviews. For each image the team must do two things: detect inappropriate or unsafe visual content before the photo is published, and read any free-form text printed on packaging visible in the photo so it can be indexed for search. There is no fixed form layout, the team has minimal machine-learning expertise, and they want to lean on managed Google Cloud AI APIs rather than train a custom model. Which TWO Google Cloud capabilities together best satisfy these two requirements? (Select TWO.)

AUse the Cloud Vision API Safe Search detection feature to flag adult, violent, or otherwise unsafe imagery before each uploaded photo is published. Correct
BUse Document AI with the invoice parser processor to extract the packaging text, since Document AI is the managed service for reading text from images.
CUse the Cloud Vision API text detection feature to extract whatever free-form text is printed on the packaging in each photo as plain strings for indexing. Correct
DUse the Cloud Translate API to convert the uploaded images into searchable text so the packaging wording can be indexed across languages.
EDeploy a custom image-classification model trained in Model Garden specifically to label each photo as safe or unsafe and to transcribe its text.

Match unstructured-image tasks to the right managed API: Vision API Safe Search screens unsafe imagery and Vision API text detection performs OCR on arbitrary photos. Both requirements operate on arbitrary, layout-free photographs, which is exactly the Vision API's domain: Safe Search returns unsafe-content likelihoods for moderation, and text detection runs OCR to extract free-form printed text. Document AI targets structured documents, Translate operates on text not images, and a custom Model Garden model is unnecessary when managed Vision features already solve both tasks.

Why A is correct: Vision API Safe Search returns likelihood scores for unsafe categories such as adult and violence, which is the purpose-built managed capability for screening user-uploaded imagery before publication.

Why B is wrong: Document AI reads text, but its parsers target structured documents with known layouts such as invoices, so it is the wrong fit for arbitrary free-form text on packaging in an unstructured photo.

Why C is correct: Vision API text detection performs OCR on arbitrary images with no fixed layout, returning the printed text as plain strings, which matches the need to read packaging text for search indexing.

Why D is wrong: Translate converts text between languages and feels relevant for a global retailer, but it accepts text not images, so it cannot read or extract the packaging wording from a photograph.

Why E is wrong: A bespoke trained model could in principle do both jobs, but it contradicts the team's minimal-ML, managed-API preference when ready-made Vision API features already cover both needs.

See more PMLE practice questions, answers explained.

Exam traps in Architecting Low-Code AI Solutions

Answers that look right on this material and are not. Each one is a distractor from a different question in the PMLE bank for this domain.

Move the conversational traffic to Vertex AI batch prediction so requests are queued and processed together, smoothing out the load during peak periods.
Why it is wrong: Batching does smooth load and lowers cost, but batch prediction is asynchronous and returns results later, so it cannot serve an interactive chat that must reply within a tight real-time latency budget.
The Vision API processes only photographs while Document AI processes only PDFs, so the correct choice is determined purely by the input file format rather than the analysis required.
Why it is wrong: Tempting because Document AI is often used on PDFs and Vision on photos, but both accept overlapping formats; the real boundary is the task, since Vision targets general image understanding and Document AI targets structured document parsing, not the file type.
Raise the temperature parameter so the model commits to an answer sooner and returns fewer retried generations per request.
Why it is wrong: Temperature only changes how random the sampling is; it does not reduce the tokens billed or shorten the prompt, so it leaves both the input-token cost and the latency of these repeated calls untouched.

Examworthy is not affiliated with or endorsed by Google Cloud. Original, blueprint-aligned practice material only.