Exercise: ground and configure a retail assistant

Original fictional scenario. No cloud account, API calls or paid services are required. Difficulty: intermediate · Estimated duration: 15–20 minutes

Lumen & Co., a fictional retailer with a pharmacy counter, is launching a Gemini-based assistant with a two-person team. Each requirement below needs a grounding choice or a generation setting.

#	Requirement	Constraint
1	Answer questions from the internal returns-policy documents	Must launch in two weeks with little engineering
2	Mention product recalls announced publicly today	The facts are public and change daily
3	The pharmacy counter needs web facts too	Customer data must not be logged
4	Returns answers must word the policy the same way every time	Runs on a model that honors custom sampling values
5	The assistant is public and used by families	Deployed on gemini-3.5-flash; nobody has touched the safety settings

Your decision

Choose a grounding option for requirements 1, 2 and 3, and say which kind of data each one is.
For requirement 4, choose the sampling change and a length control, and state when that sampling change would have no effect.
For requirement 5, state the current state of the configurable content filters and what the team must do.
Name one risk that grounding does not remove, even with the right source.

Rubric (10 house points)

3 points: correct grounding option and kind of data for each of requirements 1–3.
2 points: requirement 4 lowers temperature and caps output length, and notes Gemini 3.6 Flash and later ignore custom sampling values.
3 points: requirement 5 recognizes the filters default to OFF, sets thresholds per harm category, and layers system instructions.
2 points: names irrelevant retrieval (grounded but off-topic answers) as a risk grounding does not remove.

Reference solution

Prebuilt RAG with Agent Search over the policy documents — first-party enterprise data. Agent Search is a fully managed, out-of-the-box search and RAG builder, which suits a two-person team on a deadline.
Grounding with Google Search — world data, for up-to-date public information. Agent Search grounding can be combined with it, so the same assistant can use both.
Web Grounding for Enterprise for the pharmacy counter — still world data, but from an index suited to regulated industries, and the service doesn't log customer data.
Lower the temperature for predictable wording and set maxOutputTokens to cap length. On Gemini 3.6 Flash and later, custom temperature values are ignored, so on those models only the length control would change anything.
The configurable filters are OFF — the default for gemini-3.5-flash and later. Set blocking thresholds per harm category for a family audience, and add system instructions to steer topics, because filters block output but don't steer the model.

Risk grounding does not remove: if retrieval returns irrelevant documents, answers can be grounded but off-topic or incorrect, so retrieval quality must be tested.

Rejected alternatives. Assembling the RAG APIs for requirement 1 gives control the team does not need yet. Using Grounding with Google Search for requirement 3 does not meet the no-logging constraint. Raising temperature for requirement 4 adds variety where consistency is required. Leaving requirement 5 at the defaults assumes a protection that is switched off.

Sources and scope

The retailer, requirements and rubric are house-authored. The grounding options, settings and defaults are grounded in:

Ground and configure a retail assistant — decision exercise

Exercise: ground and configure a retail assistant

Your decision

Rubric (10 house points)

Reference solution

Sources and scope