pgvector ships two index types. Which should be your default reach in 2026, and why was it added later than the other?

Attempt it

HNSW (added in pgvector 0.5.0, mid-2023) is the default reach in 2026. It gives higher recall at the same latency than IVFFlat, has no training step, and handles ongoing upserts without periodic rebuilds.

IVFFlat is older (pgvector 0.4.0, 2022) and was the original index because it was simpler to implement inside Postgres' access-method framework: k-means + a few inverted lists. The downside is that it needs a representative training sample at build time (so an empty table cannot be indexed well) and recall under inserts degrades until the next reindex.

Reach for IVFFlat only if you specifically need its smaller memory footprint and your write pattern is batchy. For everything else in 2026, HNSW.

Your answer

Quick reference

TL;DR

HNSW is the 2026 default in pgvector. IVFFlat predates it because it was easier to bolt onto Postgres' access-method API; HNSW landed in pgvector 0.5.0 (mid-2023) and now dominates.

Memory aid

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

Easy to grasp

pgvector is like a Postgres extension that adds two filing systems for vectors. The **older one (IVFFlat)** sorts your vectors into a few hundred labeled bins, but the labels are decided when you first build the index. If you keep adding new vectors, the bins drift and your search quality slowly degrades until you rebuild. The **newer one (HNSW)** builds a graph of connections that you can keep extending as new vectors arrive, without rebuilding. Everyone reaches for HNSW first unless they specifically need the smaller memory footprint of IVFFlat.

Key concepts

pgvector

Quick scan · 2–4 min

Everything important, quickly.

2 min: name HNSW as the 2026 default + cite pgvector 0.5.0 timing + IVFFlat training/drift weakness + Postgres MVCC reason HNSW took longer.

Side-by-side

Property	IVFFlat	HNSW
pgvector version	0.4.0 (2022)	0.5.0 (mid-2023)
Algorithm family	Partition (k-means)	Graph (small-world)
Training required	Yes (representative sample)	No
Inserts degrade recall	Yes (drift over time)	No
Memory footprint	Smaller	Larger (graph + vectors)
Recall ceiling	Lower at same latency	Higher at same latency
Default reach in 2026	Memory-constrained only	Yes

Real-world examples

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

Supabase, Neon, and AWS RDS for Postgres all ship pgvector with HNSW support; their cookbook examples default to HNSW.
pgvectorscale (Timescale) adds disk-based HNSW variants for billion-scale workloads on Postgres.
Crunchy Data Postgres tutorials in 2024+ recommend HNSW as the default; IVFFlat is positioned as the memory-constrained alternative.
Hugging Face's vector tutorials for Postgres-backed RAG default to HNSW in pgvector.

Interviewer follow-ups

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

QWhat is the rule of thumb for `lists` in IVFFlat?

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

QHow does HNSW interact with Postgres MVCC?

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

QWhat does pgvectorscale add?

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

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.