AIMentor — AI Expertise Learning Platform

15 questions · pass mark 80% · 20 minute limit. The timer auto-submits at zero.

Time remaining: 20:00

Question 1easy
Why can pure vector search miss a query for an exact token like an error code 'ERR_2043'?
Vector search only works on imagesRare exact tokens carry little semantic signal, so they may not land near the right chunk in embedding spaceError codes are always filtered outCosine similarity cannot handle numbers
Question 2easy
What is hybrid search?
Searching two vector databases at onceRunning both keyword (BM25) and vector search and fusing their resultsAlternating between cosine and Euclidean distanceUsing two embedding models and averaging their vectors
Question 3medium
Why does Reciprocal Rank Fusion (RRF) use rank position instead of raw scores?
Ranks are easier to storeBM25 and cosine scores live on different, incomparable scales, so fusing by rank avoids normalizing themRaw scores are always wrongIt makes retrieval faster
Question 4medium
In RRF with score 1/(K+rank), what is the role of the constant K (commonly 60)?
It sets the number of results returnedIt dampens the gap between top ranks so no single list dominatesIt is the embedding dimensionIt filters out low-ranked documents
Question 5medium
How does a cross-encoder differ from the bi-encoder used in first-pass retrieval?
A cross-encoder embeds query and chunk separatelyA cross-encoder scores the (query, chunk) pair together in one pass, which is more accurate but slowerA cross-encoder is always fasterThere is no difference
Question 6medium
Why is a cross-encoder used as a second stage over a shortlist rather than over the whole corpus?
It cannot read more than 4 chunksScoring every (query, chunk) pair jointly is too slow at corpus scale, so you rerank only a cheap top-NIt requires the chunks to be pre-sorted alphabeticallyCross-encoders only work on shortlists by design limitation
Question 7hard
Reranking returns a poor top-4. You confirm the answer chunk was never in the 25-candidate shortlist. What should you fix?
Use a bigger reranker modelWiden or improve the first-pass retrieval — a reranker can only reorder what was retrievedLower the reranker thresholdSwitch to Euclidean distance in the reranker
Question 8medium
What is query rewriting / multi-query retrieval?
Rephrasing or expanding the user's question into one or more variants to widen retrieval coverageRewriting the retrieved chunks before generationChanging the embedding model per queryCaching previous queries
Question 9hard
In the HyDE pattern, what is embedded for retrieval?
The raw user questionA hypothetical/generated answer to the question, which often lands closer to the target chunksThe chunk metadata onlyA random vector for exploration
Question 10hard
Which are valid cautions about query rewriting? (Select all that apply)
It adds an LLM call (latency + cost) on the critical pathIt can drift from the user's original intentIt should be validated against a golden set like any changeIt always improves recall, so apply it to every query
Select all that apply.
Question 11medium
What problem does MMR (maximal marginal relevance) address?
Slow embedding generationRedundant, near-duplicate chunks crowding the top-kMissing exact-term matchesSQL injection in the vector query
Question 12hard
In MMR's score λ·relevance − (1−λ)·redundancy, what does setting λ = 1.0 do?
Pure diversity, ignoring relevancePure relevance, ignoring diversity (equivalent to plain similarity ranking)Disables retrievalBalances relevance and diversity equally
Question 13medium
Which single addition is often the highest-ROI upgrade to a basic RAG retrieval pipeline?
Switching embedding providersAdding a cross-encoder reranker over a shortlistIncreasing temperatureDoubling the chunk size
Question 14medium
What is the classic keyword/term-frequency ranking algorithm (acronym) paired with vector search in hybrid retrieval?
Question 15hard
Name the rank-based method (acronym or full name) for combining multiple ranked result lists without normalizing scores.

0 / 15 answered

Retrieval Quality — Final Exam