Task 3.3: Identify grounding techniques and their use cases.

• AGrounding with Agent Search over the company's documents

  Incorrect

  Agent Search grounding is not a source of breaking news from across the web.

• BSupervised fine-tuning on last year's articles

  Incorrect

  Tuning on old articles cannot teach the model about events from the past few hours.

• CGrounding with Google Search

  Correct

  Grounding with Google Search connects the model to world knowledge and up-to-date information on the internet.

• DGrounding with Google Maps

  Incorrect

  Maps grounding adds geospatial and place information; it is not a source of general news.

Fresh public information is world data, and Grounding with Google Search is the option that brings it in, with links to the sources.

• AGrounding with Google Search

  Incorrect

  Google Search grounding cannot see the company's internal HR policies.

• BGrounding with Agent Search

  Correct

  Grounding with Agent Search connects the model to your website data or document sets stored in Agent Search, so answers are tethered to the company's own policies.

• CWeb Grounding for Enterprise

  Incorrect

  Web Grounding for Enterprise is a compliance-controlled web index for regulated industries; it is still web data, not the company's own documents.

• DRaising the temperature for broader answers

  Incorrect

  Temperature controls randomness in word choice; it does not connect the model to any source.

First-party enterprise data calls for grounding in your own data store; web grounding options cover public information.

• ATop-P (nucleus sampling)

  Incorrect

  Top-P is a sampling setting that shapes how varied the output is; it does not limit length.

• BTemperature

  Incorrect

  Temperature controls how random the token choices are; a low temperature can still produce a long answer.

• CSafety settings

  Incorrect

  Safety settings set blocking thresholds for harmful content categories; they do not control length.

• DMaximum output tokens

  Correct

  maxOutputTokens limits the number of tokens generated in the response; a token is about four characters, so 100 tokens are roughly 60 to 80 words.

Length is controlled by maximum output tokens (and by asking for a length in the prompt); randomness by temperature and top-P; harm by safety settings.

• ARAG retrains the model, so the change needs another training run

  Incorrect

  RAG does not change the model's weights; it works with the model as it is.

• BRAG works only with structured data such as database tables

  Incorrect

  RAG corpora are commonly built from documents such as manuals, split into chunks and embedded for search.

• CThe answer can only be as good as what the retrieval step finds

  Correct

  RAG can only ground an answer in documents that are in the corpus. If the relevant manual is missing, nothing useful is retrieved, and the model falls back on what it already knows.

• DRAG removes the need to evaluate the assistant's answers

  Incorrect

  RAG reduces hallucinations, but answers still need evaluation, as this gap shows.

RAG improves grounding only for content it can retrieve. Keeping the corpus complete and current is part of running a RAG system.

• AGrounding with Google Maps

  Correct

  Grounding with Google Maps gives the model Google Maps data, so responses are more accurate and context-aware, including geospatial context.

• BGrounding with Agent Search over brochures

  Incorrect

  Brochures may be outdated and incomplete about the neighborhood; the need is place data.

• CWeb Grounding for Enterprise

  Incorrect

  Web Grounding for Enterprise is a compliance-controlled web index for regulated industries, not a source of place details.

• DRAG Engine over last year's guidebooks

  Incorrect

  Old guidebooks would ground answers in stale place information.

Pick the grounding source by the kind of truth you need: places from Maps, the web from Search, your content from Agent Search or RAG Engine.

• AHigh, close to 2, for more varied wording

  Incorrect

  High temperatures make word choices more random and diverse, the opposite of what the team wants.

• BAt 1.0, the recommended starting value

  Incorrect

  1.0 is Google's recommended starting value for general use, but it still leaves room for random variation; the team wants as little as possible.

• CModerate, around 0.5, for balanced output

  Incorrect

  A middle value reduces randomness but does not minimize it; lower temperatures are more predictable still.

• DLow, close to 0, for focused output

  Correct

  Lower temperatures give more predictable responses suited to less open-ended tasks; at 0 the most probable tokens are always selected, though small variation is still possible.

Low temperature for consistency, higher for creativity. Note that the newest Gemini models ignore custom temperature, top-P and top-K values, so check the model before relying on them.

• AGrounding with Google Search

  Incorrect

  Google Search grounding offers the broadest and freshest index, but Google collects and stores logs for it; it lacks the extra compliance controls.

• BWeb Grounding for Enterprise

  Correct

  Web Grounding for Enterprise uses a web index suited to regulated industries; the service does not log customer data and supports VPC Service Controls.

• CGrounding with Google Maps

  Incorrect

  Maps grounding adds place information, not general web research.

• DFine-tuning on downloaded web pages

  Incorrect

  Tuning on a snapshot is neither current nor a grounding service.

Both web options ground in the web. Choose Web Grounding for Enterprise when compliance controls matter more than the broadest, freshest index.

• ASet dangerous content to BLOCK_ONLY_HIGH

  Correct

  Content filter thresholds are set per harm category; BLOCK_ONLY_HIGH blocks only content scored high for probability or severity, so fewer benign medical answers are blocked.

• BLower the temperature for every request

  Incorrect

  Temperature changes randomness in the answer, not whether the filter blocks it.

• CRaise the maximum output tokens

  Incorrect

  Longer outputs do not change how the filter scores them.

• DAdd a few-shot example of a blocked answer

  Incorrect

  Filters act on the response independently of the prompt's examples; they act as a barrier, not as model guidance.

Safety settings trade false positives against false negatives. Tune the threshold for the specific category that misfires, and test the result.

• ATop-K

  Incorrect

  Top-K keeps a fixed number of the most probable tokens; it does not use a cumulative probability threshold, so the set size cannot vary.

• BTemperature

  Incorrect

  Temperature sets how randomly the final token is picked from the candidates; it does not decide which candidates are kept.

• CFrequency penalty

  Incorrect

  A frequency penalty lowers the probability of tokens that already repeat in the output; it does not cut the candidates at a probability total.

• DTop-P (nucleus sampling)

  Correct

  With Top-P, tokens are selected from most to least probable until the sum of their probabilities equals the Top-P value; less likely tokens are excluded.

Top-K fixes how many candidates are kept, Top-P fixes their combined probability, and temperature then sets how randomly the pick is made.