Vector databases and similarity search quiz

Question 1

What is the core trade-off that Approximate Nearest Neighbor (ANN) algorithms make compared to exact k-NN search?

Accepted Answer

Recall for lower latency and better scalability

Answer

Memory for higher precision

Answer

Index build time for lower recall

Answer

Dimensionality for exactness

Question 2

In HNSW, what is the role of the multi-layer graph structure?

Accepted Answer

Upper layers give long-range express links; lower layers refine locally

Answer

All layers store identical neighbor sets for redundancy

Answer

Layers partition vectors by cluster centroid

Answer

Each layer stores a different quantization codebook

Question 3

Why does cosine similarity often outperform Euclidean (L2) distance for text embeddings?

Accepted Answer

It captures direction/semantics and ignores magnitude differences

Answer

It is always faster to compute than L2

Answer

It accounts for vector sparsity automatically

Answer

It guarantees a metric space unlike L2

Question 4

What does Product Quantization (PQ) do, and why is it useful in vector databases?

Accepted Answer

Compresses vectors into sub-vector codebook IDs for memory-efficient approximate distances

Answer

Reduces dimensionality via PCA before indexing

Answer

Replaces the distance metric with Hamming distance

Answer

Partitions the dataset into disjoint shards

Question 5

In an IVF (Inverted File) index, what do the parameters nlist and nprobe control?

Accepted Answer

nlist = number of clusters; nprobe = number of clusters searched per query

Answer

nlist = vector dimension; nprobe = number of probes per second

Answer

nlist = list length per cluster; nprobe = quantizer bits

Answer

nlist = number of layers; nprobe = entry points

Question 6

What is the curse of dimensionality and how does it affect similarity search?

Accepted Answer

Distances concentrate, reducing neighbor contrast and degrading tree indexes

Answer

Memory usage grows quadratically with dimensions

Answer

Cosine similarity becomes undefined above 1000 dims

Answer

Vectors automatically become orthogonal and identical

Question 7

What are the two main HNSW build/search parameters and their effects?

Accepted Answer

M sets links per node; ef controls candidate list size for build/search

Answer

M sets number of layers; ef sets the metric

Answer

M is the metric; ef is the shard count

Answer

M is memory in MB; ef is epochs

Question 8

What is filtered (metadata) vector search, and why is pre-filtering vs post-filtering a key design decision?

Accepted Answer

Combining vectors with metadata filters; pre vs post affects recall and completeness

Answer

A way to remove duplicate vectors before indexing

Answer

A method to normalize vectors per category

Answer

Filtering out low-magnitude vectors automatically

Question 9

How does hybrid search typically combine dense (vector) and sparse (lexical/BM25) retrieval?

Accepted Answer

Fuse dense and sparse rankings, often via Reciprocal Rank Fusion

Answer

Average the embeddings of both methods

Answer

Use BM25 only when vectors fail

Answer

Concatenate dense and sparse vectors before indexing

Question 10

What is the relationship between dot product, cosine similarity, and Euclidean distance for normalized vectors?

Accepted Answer

For unit vectors all three give identical rankings (L2^2 = 2 - 2cos theta)

Answer

They always disagree on rankings

Answer

Dot product equals L2 distance for unit vectors

Answer

Cosine is undefined for normalized vectors

Question 11

What is recall@k as an evaluation metric for ANN indexes, and how is it measured?

Accepted Answer

Overlap of ANN top-k with exact top-k, averaged over queries

Answer

Precision of the embedding model

Answer

Fraction of queries returning any result

Answer

Number of vectors indexed per second

Question 12

Why is Maximum Inner Product Search (MIPS) not a true metric space, and what's a common workaround?

Accepted Answer

IP violates triangle inequality; augment dimensions to reduce MIPS to L2 NN

Answer

IP is always negative so it must be squared

Answer

MIPS requires normalization to become a metric, which removes all info

Answer

There is no workaround; MIPS cannot be indexed

Question 13

What is the purpose of a re-ranking (rescoring) stage after ANN retrieval in a RAG pipeline?

Accepted Answer

Rescore ANN candidates with a more accurate model (e.g., cross-encoder)

Answer

Compress the candidate vectors further

Answer

Deduplicate the embeddings

Answer

Convert dense vectors to sparse

Question 14

What distinguishes a bi-encoder from a cross-encoder, and why are both used together?

Accepted Answer

Bi-encoder encodes separately (fast retrieval); cross-encoder jointly (accurate rerank)

Answer

Bi-encoder is for images, cross-encoder for text

Answer

They are identical but cross-encoder uses GPUs

Answer

Cross-encoder produces vectors; bi-encoder produces scores

Question 15

What is IVF-PQ and why is it favored for billion-scale similarity search?

Accepted Answer

IVF clusters + PQ compression for memory-efficient billion-scale ANN

Answer

Two HNSW graphs combined

Answer

Inverted full-text index with PageRank

Answer

Pure brute-force on GPU

Question 16

What is the difference between sharding and replication in a distributed vector database?

Accepted Answer

Sharding splits data for capacity; replication copies it for availability/QPS

Answer

Sharding copies data; replication splits it

Answer

Both mean the same compression scheme

Answer

Sharding is for vectors, replication for metadata only

Question 17

How does Scalar Quantization (SQ) differ from Product Quantization (PQ)?

Accepted Answer

SQ quantizes each dimension to int (~4x, simple); PQ quantizes sub-vectors via codebooks (higher compression)

Answer

SQ and PQ are identical

Answer

SQ requires k-means training; PQ does not

Answer

SQ increases dimensions; PQ reduces them

Question 18

What is a common cause of poor recall when adding a metadata filter to an HNSW search, and a mitigation?

Accepted Answer

Selective post-filter drops results; mitigate with over-fetch or filter-aware traversal

Answer

HNSW cannot use metadata at all

Answer

Filters always improve recall

Answer

Increase M to fix filtering

Question 19

What is the role of the embedding model's output dimensionality in vector DB cost and performance?

Accepted Answer

Higher dims raise storage/latency; Matryoshka enables truncation trade-offs

Answer

Dimensionality has no effect on cost

Answer

Lower dims always improve accuracy

Answer

Only sparse vectors are affected by dimensionality

Question 20

What is Reciprocal Rank Fusion (RRF) and why is it popular for fusing retrieval results?

Accepted Answer

Rank-based fusion summing 1/(k+rank); needs no score normalization

Answer

Averages raw similarity scores

Answer

Trains a neural fusion layer

Answer

Picks the single best-ranked list

Question 21

Why might you choose a disk-based ANN index like DiskANN over an in-memory HNSW index?

Accepted Answer

To serve billion-scale indexes cheaply from SSD with limited RAM

Answer

Because SSD is always faster than RAM

Answer

Disk indexes give exact results unlike HNSW

Answer

HNSW cannot handle more than 1M vectors

Question 22

What is the impact of chunking strategy on retrieval quality in a vector-search RAG system?

Accepted Answer

Chunk size/overlap balances embedding precision vs retained context

Answer

Chunking only affects storage cost

Answer

Smaller chunks always improve quality

Answer

Chunking is irrelevant once embeddings exist

Question 23

How do you handle updates and deletions in graph-based ANN indexes like HNSW?

Accepted Answer

Use soft deletes/tombstones with periodic rebuild; hard deletes are problematic

Answer

Deletions are O(1) and trivial in HNSW

Answer

You must rebuild the whole index for every insert

Answer

HNSW does not support inserts after build

Question 24

What does the 'over-retrieval then re-rank' pattern fix, and what is its main cost?

Accepted Answer

Fixes ANN imprecision at the cost of extra rerank latency/compute

Answer

Reduces storage at the cost of recall

Answer

Eliminates the need for embeddings

Answer

Removes the need for an ANN index

Question 25

What is Locality-Sensitive Hashing (LSH) and where does it fit relative to modern ANN methods?

Accepted Answer

Hash-based bucketing of near points; largely outperformed by HNSW/IVF-PQ today

Answer

The default index in all modern vector DBs

Answer

A lossless compression method

Answer

Exact search via cryptographic hashes

Question 26

How does GPU acceleration help vector similarity search, and what are its limits?

Accepted Answer

Great for batched/brute-force and training; limited by VRAM and irregular graph traversal

Answer

GPUs make HNSW single queries always fastest

Answer

GPUs eliminate the need for any index

Answer

GPUs only help with metadata filtering

Question 27

Why is L2-normalizing embeddings before indexing a common best practice when using inner-product indexes?

Accepted Answer

It makes inner product equal cosine, giving direction-based semantics

Answer

It reduces vector dimensionality

Answer

It is required for metadata filtering

Answer

It compresses vectors to int8

Question 28

What is the difference between latency, throughput (QPS), and recall when benchmarking a vector index?

Accepted Answer

Per-query time vs queries/sec vs accuracy; they trade off on a Pareto curve

Answer

They are three names for the same metric

Answer

Throughput always equals 1/latency exactly

Answer

Recall is independent of efSearch/nprobe

Question 29

What problem does Matryoshka Representation Learning (MRL) solve for vector databases?

Accepted Answer

Embeddings truncatable to shorter prefixes with graceful accuracy loss

Answer

A new distance metric for sparse vectors

Answer

A way to encrypt embeddings

Answer

A graph index replacing HNSW

Question 30

In FAISS, what is the difference between IndexFlatL2 and an IVF/HNSW index?

Accepted Answer

Flat is exact brute-force baseline; IVF/HNSW are approximate and scalable

Answer

Flat is approximate; IVF is exact

Answer

All three are exact

Answer

Flat uses a graph; IVF uses brute force

Test yourself