Try each one before opening the answer. Every option is explained, with the Databricks documentation page that proves it.
Question 1 · choose 1
A team wants to try hybrid search and the reranker on its AI Search index. It has no labeled query-answer pairs and no time to build a large golden set. Following the AI Search retrieval quality guide, how should it proceed?
ABuild a small synthetic eval set and compare relative gains
BHold off on any tuning until a large labeled golden dataset exists
CTurn on both features and judge them from a few manual test queries
DCompare the index's absolute scores with a published industry benchmark
Show the answer and why
ABuild a small synthetic eval set and compare relative gains
Correct
The guide asks for some automated measurement first; even a small synthetic set shows whether a technique helps, and relative improvements matter more than absolute scores.
BHold off on any tuning until a large labeled golden dataset exists
Incorrect
Perfect data is not required; a golden set is one option next to synthetic sets and ground-truth-free judges.
CTurn on both features and judge them from a few manual test queries
Incorrect
Without a reproducible evaluation, optimizing is guesswork, and a handful of manual checks is not a measurement.
DCompare the index's absolute scores with a published industry benchmark
Incorrect
Scores on other data say little about this index; the guide focuses on relative improvements between strategies on your own data.
Before optimizing retrieval, establish an automated evaluation: an existing golden dataset, a synthetic set generated from the documents, or ground-truth-free judges. Then test each technique against it.
Two applications share an AI Search index and both rank poorly: a RAG agent at about 3 QPS with a 2-second time-to-first-token target, and a product search box at 150 QPS that must answer in under 100 ms. Which reranking plan fits both?
AUse the built-in reranker for both applications, with the same columns_to_rerank for each
BBuilt-in reranker for the agent; a light cross-encoder on Model Serving for search
CUse the built-in reranker for the search box, and raise num_results for the RAG agent
DSkip reranking and switch both applications to full-text queries
Show the answer and why
AUse the built-in reranker for both applications, with the same columns_to_rerank for each
Incorrect
The built-in reranker handles about 5 QPS out of the box and typically takes hundreds of milliseconds, which breaks the search box's limits.
BBuilt-in reranker for the agent; a light cross-encoder on Model Serving for search
Correct
The built-in reranker suits low-QPS RAG agents where generation dominates latency; sub-100 ms, high-QPS search needs a faster reranker served as a custom model.
CUse the built-in reranker for the search box, and raise num_results for the RAG agent
Incorrect
This puts the slower reranker on the latency-critical path and gives the agent more noise instead of better ordering.
DSkip reranking and switch both applications to full-text queries
Incorrect
Full-text search matches exact terms; it does not reorder semantically relevant results and is not a reranker.
The AI Search reranker improves quality (roughly 10 to 15 percent) at a latency cost, which suits RAG agents and low-to-moderate QPS. For search bars that need under 100 ms, Databricks suggests a lightweight reranking model on Model Serving.
During an incident, AI Search queries that use the reranker still return results, with a RERANKER_TEMPORARILY_UNAVAILABLE warning in the debug info. What does this mean for those results?
AThey were returned without reranking
BThey are empty placeholders for the hits
CThey came from a cached earlier response
DThey were reranked twice by a fallback
Show the answer and why
AThey were returned without reranking
Correct
When the reranker call fails, the results returned have not been processed by the reranker.
BThey are empty placeholders for the hits
Incorrect
Results are returned; only the reranking step was skipped.
CThey came from a cached earlier response
Incorrect
The warning does not refer to caching.
DThey were reranked twice by a fallback
Incorrect
The warning means reranking did not run at all.
AI Search degrades gracefully: the retrieval results are returned without reranking, and the debug message carries the warning so applications can log or retry later.