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 does the plain Python-list index stop working as your corpus grows large?
Python lists can't store floatsIt scans and re-scores every stored vector per query — O(N) — which is hopeless at millions of vectorsCosine similarity is inaccurate above 1000 vectorsLists lose precision over time
Question 2easy
What does approximate nearest-neighbor (ANN) search trade for its speedup?
Memory for latencyA small amount of accuracy (recall) for a large reduction in vectors searchedEmbedding dimensions for storageNothing — it is exact and fast
Question 3medium
recall@k in ANN search measures:
How fast the query returnsThe fraction of the true top-k neighbors the index actually returnedThe number of dimensions in each vectorHow much memory the index uses
Question 4medium
For general text-embedding retrieval, which similarity metric is the safe default?
Euclidean (L2)Cosine similarityJaccard indexHamming distance
Question 5hard
If all vectors are normalized to unit length, how do cosine, dot product, and L2 rank results?
Completely differentlyIdentically — normalization makes them produce the same rankingOnly cosine and L2 agreeRanking becomes random
Question 6medium
In pgvector, what does the `<=>` operator compute?
Cosine distance (1 − cosine similarity)Euclidean distanceString equalityNegative inner product
Question 7medium
Why use %s parameterized placeholders instead of building the SQL string with the values?
It runs fasterThe driver safely substitutes values (prevents SQL injection and quoting bugs)pgvector requires it for vectors onlyIt is purely a style preference
Question 8medium
What is the advantage of filtering by metadata inside the vector DB (a WHERE clause) versus filtering in Python after retrieval?
It changes the embedding modelThe engine narrows the candidate set efficiently at scale instead of you post-filtering resultsIt removes the need for an ANN indexThere is no difference
Question 9hard
Why must you specifically check recall on FILTERED ANN queries?
Filters always improve recallA very selective filter can leave the ANN traversal returning too few matching rows (post-filter starvation)Filtered queries can't use vectorsWHERE clauses disable the index permanently
Question 10hard
In pgvector's HNSW index, which parameter is the per-query recall/latency knob?
mef_constructionef_searchvector_cosine_ops
Question 11hard
Which statements about HNSW vs. IVF indexes are correct? (Select all that apply)
HNSW generally offers better recall at a given latencyIVF (IVFFlat) typically builds faster and uses less memoryIVF's query knob is the number of probes (lists searched)HNSW requires no index build at all
Select all that apply.
Question 12medium
A team already runs Postgres and has ~300k chunks. What is the pragmatic default vector store?
pgvectorA dedicated Qdrant clusterRebuild everything on PineconeKeep the Python list
Question 13medium
On Azure, which store is the default for enterprise RAG needing hybrid (keyword + vector) search and semantic reranking?
pgvector on a VMAzure AI SearchA local FAISS fileBlob storage
Question 14medium
Name the default graph-based ANN index (acronym) exposed by pgvector and most vector DBs.
Question 15hard
cosine distance equals 1 minus what quantity? (two words)

0 / 15 answered

Vector Databases — Final Exam