Week 19: Quiz (10 questions, 8/10 to pass)

1. MCQ: Google AI Studio and Vertex AI expose largely the same Gemini models. What is the single most important difference for sensitive data? (see Concepts §AI Studio vs Vertex)

   • A) Vertex models are faster
   • B) AI Studio's free tier may use your data to improve Google's products by default
   • C) Vertex has no quotas
   • D) AI Studio requires a cloud project

2. MCQ: In the google-genai SDK, what flips the same client from AI Studio to Vertex? (see notebook 02 cell 8)

   • A) Changing the model name
   • B) Setting vertexai=True plus project/location
   • C) Using a different import
   • D) Setting response_mime_type

3. MCQ: Which Gemini rung is the cheapest for high-volume, simple extraction? (see Concepts §Gemini ladder)

   • A) gemini-2.5-pro
   • B) gemini-2.5-flash
   • C) gemini-2.5-flash-lite
   • D) A "Thinking" variant

4. Short answer: In notebook 01, why does the pipeline render each BoL text to a PNG with Pillow before calling Gemini, instead of just pasting the text into the prompt? (see notebook 01 cell 3)

5. MCQ: A 429 error mid-batch on Vertex is almost always… (see Concepts §How it breaks)

   • A) A syntax bug
   • B) A quota ceiling, raise per-region/per-model quotas or spread calls out
   • C) A wrong API key
   • D) A model that doesn't exist

6. MCQ: Which of Google's three agent layers is the code-first, open-source framework you rebuild the support agent with? (see Concepts §Agent stack)

   • A) Agent Builder
   • B) Agent Engine
   • C) ADK
   • D) Colab Enterprise

7. Short answer: Name at least five of the ten BoL fields the OCR prompt returns as JSON. (see notebook 01 cell 6)

8. MCQ: ML.GENERATE_TEXT in BigQuery ML lets you… (see Concepts §BigQuery ML)

   • A) Train a transformer from scratch in SQL
   • B) Call Gemini for generation/classification from inside BigQuery without leaving SQL
   • C) Replace Vertex IAM
   • D) Render images

9. Short answer: Using the notebook's placeholder prices ($0.000315/image, $0.30/1M input, $2.50/1M output), compute the cost of one BoL document that is one image, 150 input tokens, and 120 output tokens. Show the arithmetic. (see notebook 01 cell 11)

10. MCQ: "A pipeline without a per-field number is just a screenshot of a JSON blob" means… (see Concepts §How it breaks)

    • A) JSON output is never correct
    • B) You must measure accuracy field-by-field against ground truth, not just eyeball a pretty blob
    • C) Screenshots are the deliverable
    • D) Only total accuracy matters, per-field is optional

Answer key

1. B: AI Studio's free tier may use data to improve Google's products by default; Vertex is where governed data goes. The segregation habit is the deliverable.
2. B: genai.Client(vertexai=True, project=..., location=...) switches auth and billing to Vertex with ADC.
3. C: flash-lite is the cheapest rung for high-volume/simple tasks; pick the cheapest model that clears your accuracy bar.
4. To make Gemini exercise vision, not read text from the prompt, rendering to an image keeps the pipeline end-to-end (image → JSON) and tests the OCR capability that is Google's edge.
5. B: 429 is a quota problem, not a code bug; check and raise per-region/per-model quotas before the batch.
6. C: ADK is the code-first open-source framework; Agent Builder is low-code, Agent Engine is the managed runtime.
7. Any five of: shipper, consignee, port_of_loading, port_of_discharge, commodity, quantity, gross_weight_kg, declared_value_usd, freight_terms, date_of_issue.
8. B: ML.GENERATE_TEXT calls a remote Gemini model from inside BigQuery SQL, "AI without leaving the warehouse."
9. Image $0.000315 + input 150 × $0.30/1M = $0.000045 + output 120 × $2.50/1M = $0.00030 → total ≈ $0.00066.
10. B: A correct-looking JSON blob can still have the wrong consignee; per-field accuracy against ground truth is the only honest metric.