FDEBench Join us live

Q
6 / 159

Grounding & safety

What are the key steps to design a system that allows Gemini to browse the web in real time and provide accurate answers without hallucinating outdated facts?

question bank FDEBench bank Source

01
Ask

What a strong candidate asks first

Before designing anything. Skipping this is the most common way the round is lost.

  • What's the latency budget? Live fetching is slow.
  • Is "I don't know" an acceptable answer, or must it always respond?
  • Are we answering from the model's knowledge when retrieval fails?

02
Rubric

What earns the points

7 criteria, 100 points total. Scored independently, partial credit allowed.

Browses only when it's needed 15 pts

Decides per-query whether live retrieval is needed rather than fetching on every request, and says how that decision is made.

If you miss it You fetched the web on every query. Most questions don't need live data — routing only the time-sensitive ones saves the latency and the cost.
Rewrites questions into searches 10 pts

Transforms the user's question into search queries rather than passing it through verbatim.

If you miss it You searched with the user's raw question. Conversational phrasing makes a poor search query — rewrite it into targeted queries first.
Answers only from what it found 25 pts

Constrains generation to the retrieved content and says what happens when retrieval comes back empty or thin — refuse, or clearly mark model knowledge as unverified. The empty-retrieval branch is required for full credit.

If you miss it You grounded the answer but never said what happens when retrieval returns nothing useful. That's exactly the case where the model falls back on stale training data and states it confidently.
Ranks sources by trust 15 pts

Distinguishes sources by authority or trust, not just relevance and recency. A recent low-quality page is not better than an older authoritative one.

If you miss it You ranked on relevance and recency only. The freshest result is often the worst one — source authority has to be part of the ranking.
Flags when sources disagree 15 pts

Detects when retrieved sources disagree and surfaces the disagreement rather than silently picking one.

If you miss it When your sources conflict, your design picks one arbitrarily. Detect disagreement and say so — that's more useful than a confident wrong answer.
Cites per claim and checks it 10 pts

Attaches citations at the claim level, and checks they actually support the claim, rather than appending a link list.

If you miss it Your citations are a list at the bottom. Attach them to specific claims and verify the source actually supports the sentence — an unverified citation is worse than none.
Spots stale and dead pages 10 pts

Handles pages that are stale or deleted — publication dates, cache age, or detecting a dead URL.

If you miss it You trusted whatever the fetch returned. Pages get deleted and rewritten; check publication dates and handle dead or changed URLs.

03
Flags

Red flags

Reported separately, not subtracted — this is the "what would have ended the interview" note.

  • Falls back to model knowledge on retrieval failure without flagging it.
  • No latency consideration for live fetching.

—