Product design question

How would you improve Glean's enterprise search relevance across 100+ connectors?

Practice this question out loud. An AI interviewer asks it, follows up like a real interviewer would, and scores your answer. Type or speak.

What this question tests

Enterprise search relevance design at scale, requiring judgment about ranking signals across fragmented data sources.

Clarify the challenge: relevance across 100-plus connectors (Slack, Google Drive, Confluence, Jira, and more) means content varies wildly in structure, freshness, and trust level.
Identify current likely weaknesses: recency bias missing older but still-relevant documents, and permission-aware ranking that may surface technically-accessible but low-relevance content.
Propose a source-quality signal: weight results by document freshness, author authority (e.g., an official wiki page over a stray chat message), and historical click-through for similar queries.
Propose a personalization layer: rank results higher if they come from the querying user's own team, frequently used tools, or recent work context.
Add explicit feedback loops: let users mark a result as helpful or not, feeding back into ranking, since enterprise search lacks the scale of consumer search click data.
Define success metrics: click-through rate on top-3 results, and a qualitative satisfaction score gathered via periodic surveys given limited natural feedback signal in enterprise settings.

Names the specific challenge of enterprise search: fragmented, heterogeneous sources with far less natural feedback signal than web search.
Proposes concrete ranking signals (freshness, source authority, personalization) tied to enterprise data realities.
Addresses the cold-start-like problem of sparse click data with an explicit feedback mechanism.
Balances personalization against permission constraints, since relevance ranking must respect access control.
Uses both quantitative (click-through) and qualitative (survey) metrics given enterprise search's smaller data volume.

Proposing a generic web-search-style ranking algorithm without addressing enterprise-specific data fragmentation.
Ignoring permission constraints when discussing ranking and personalization.

How would you handle a source with poor or missing metadata?
How would permissions interact with your ranking signals?
How would you gather enough feedback signal in a lower-traffic enterprise setting?