Vector Databases & ANN Search — Advanced

HNSW/IVF/PQ internals, OPQ and ScaNN, scalar/binary quantization + rescoring, filtered ANN, disk-based indexes, MIPS, and distributed sharded search.

Hard50 questions

Sample questions

HNSW search keeps a candidate list of size 'ef' per layer. What does that beam do?

It caps the candidates the greedy search tracks
It stores the metadata payloads for filtering
It sets the number of graph layers built
It encrypts the visited nodes

Why

ef caps the size of the dynamic candidate set the search maintains while hopping the graph; a larger ef explores more neighbors before stopping, raising recall at higher latency. It isn't payload storage, the layer count, or encryption.

HNSW paper Pinecone: HNSW

How does HNSW decide which layer a newly inserted vector reaches?

A random level, exponentially decaying by layer
By the vector's L2 norm
By its insertion timestamp
By its cosine similarity to the origin

Why

Each node's maximum layer is sampled so that higher layers get exponentially fewer nodes, creating the sparse long-range top and dense bottom that make search efficient. It's not based on norm, timestamp, or similarity to the origin.

HNSW paper Pinecone: HNSW

Why are deletions notoriously awkward in an HNSW graph index?

HNSW cannot store more than a million vectors
Removing a node breaks connectivity; use tombstones
Deletion doubles the index memory
Deleted vectors still appear in every result

Why

A node may be an important hub linking regions, so truly removing it risks disconnecting the graph; systems typically mark it deleted (tombstone) and periodically rebuild/compact. HNSW scales past millions, deletion doesn't double memory, and tombstoned nodes are filtered out.

HNSW paper Qdrant: What is a vector database

In Product Quantization, a vector is split into m sub-vectors, each quantized with a codebook of 2^b centroids. How many bytes encode one vector (b=8)?

2^b bytes regardless of m
One byte for the whole vector
m bytes, one code per sub-vector
The original dimension in bytes

Why

With b=8, each sub-vector's nearest centroid id is one byte, so m sub-vectors take m bytes total — a large compression versus the raw float vector. It's not 2^b bytes, a single byte, or the original size.

FAISS: Billion-scale similarity search Pinecone: Product Quantization

PQ search typically uses 'asymmetric distance computation' (ADC). What is asymmetric about it?

The query stays full-precision while database vectors are quantized
The query is quantized but database vectors are not
Only half the dimensions are compared
The distance is computed twice and averaged

Why

ADC keeps the query in full precision and precomputes its distance to each sub-codebook centroid, then looks up distances to the quantized database codes — more accurate than quantizing the query too. It's not query-only quantization, half-dimension, or double computation.

FAISS: Billion-scale similarity search Pinecone: Product Quantization

Free account

Take the full module

These are the first few of 50 questions. A free account opens the rest as a scored drill.

Every question in this module
Instant feedback and supporting reading
Your score and progress, saved

Free · your email is used for progress only.

About Nybble^™

The AI space moves fast.
Nybble^™ is how you keep up — and stay sharp.

Snack

What happened. In two minutes.

The AI news cycle moves at a pace no one can keep up with. Snack distills what launched, what shipped, and what matters — every day, without the filler.

Go to Snack

Stack

The concepts behind the headlines.

News tells you what. Stack tells you why and how. From RAG architectures to agentic evals, these are the ideas that will shape what you build next.

Go to Stack

Hack

Prove you actually get it.

Reading about LangChain is not the same as knowing it. Hack challenges you with production-grade questions, then shows you the references that make the answer stick.

Go to Hack