Fill in the IVF nlist rule of thumb.

Attempt it

For an IVF index over N vectors, a common rule of thumb is to set nlist (the number of clusters) to approximately , because going much larger makes the per query centroid comparison step dominate, and going much smaller forces each partition to scan a large slice.

Quick reference

TL;DR

nlist ~ sqrt(N). Larger makes centroid comparison dominate per query work; smaller forces each partition to be large and slow to scan.

Memory aid

Sign in for the mnemonic that makes this stick. See Pro pricing.

Easy to grasp

Imagine sorting a million books into boxes. Too few boxes means each box is huge and you waste time digging through one. Too many boxes means you spend forever picking which box to look in. The sweet spot is when the number of boxes is about the same as the number of books per box, which math says is roughly sqrt(N).

Key concepts

IVF (Inverted File) index K-means clustering

Detailed answer & concept explanation

~2 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.

Sign in for 5 free unlocks/day or go Pro for unlimited. See Pro pricing.

Quick scan · 2–4 min

Everything important, quickly.

Derive the optimum from T(nlist) = nlist + nprobe * N / nlist. Discuss the 4*sqrt(N) refinement, k-means training cost at scale, and why nprobe (not nlist) is the recall knob.

Real-world examples

Real products, models, and research that use this idea.

FAISS wiki - Guidelines to choose an index lists nlist = sqrt(N) to 16*sqrt(N) depending on recall target and corpus size.
Milvus IVF index docs recommend nlist starting at 4*sqrt(N) for production, often refined upward.
Pinecone p2 pods use IVF-PQ internally with nlist tuned around 4*sqrt(N) at their billion-vector scale.
OpenSearch's IVF implementation defaults to nlist = sqrt(N) and exposes it as a build-time parameter.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QWhy does the per query work split evenly when nlist = sqrt(N)?

Sign in or go Pro to see the approach for this follow-up.

QWhat is the difference between nlist and nprobe?

Sign in or go Pro to see the approach for this follow-up.

QHow do you know when to use 4*sqrt(N) vs sqrt(N)?

Sign in or go Pro to see the approach for this follow-up.

Don't say this

Red flags and common mistakes that signal junior thinking. Click to expand.

Most common mistake

Setting nlist proportional to N (linear growth) or treating it as a recall knob; the recall knob is nprobe, and nlist is a structural balance.

Sign in or go Pro to see all red flags and common mistakes. See Pro pricing.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

Sign in for 5 unlocks/day or go Pro for unlimited revision sheets. See Pro pricing.

References

Primary sources. Skim if you want the original framing.