Dev Tools

PGVector Is Still The Right Answer In 2026

All articles
🐘📊🛠️

Postgres you already run, plus a CREATE EXTENSION

If you've been living under a rock of vector-DB hype, here's the unsexy truth: pgvector inside the Postgres you already run handles 95% of real-world workloads, and the half-precision quants make it embarrassingly competitive at scale.

The Setup

We use this exact pattern at Rebuild Relief — claims, policy docs, and Pipedrive notes all live in Supabase Postgres with pgvector tacked on. Zero new infrastructure, same backup story, same RLS policies, same SQL my team already knows.

-- enable the extension (already on Supabase by default)
create extension if not exists vector;

-- halfvec = 16-bit floats, half the storage, ~zero recall loss
create table claim_embeddings (
  id          bigint primary key,
  claim_id    uuid not null references claims(id) on delete cascade,
  chunk       text not null,
  embedding   halfvec(1024) not null
);

-- HNSW index for fast ANN
create index on claim_embeddings
  using hnsw (embedding halfvec_cosine_ops)
  with (m = 16, ef_construction = 64);

The Money Pattern

Query is one SQL statement. Joins, filters, RLS, all the Postgres machinery you already know — and the vector distance is just another operator. Try doing this on Pinecone without writing a sidecar service.

-- top 5 most similar chunks, filtered by region, respecting RLS
select
  c.claim_id,
  c.chunk,
  c.embedding <=> $1::halfvec as distance,
  cl.region,
  cl.status
from claim_embeddings c
join claims cl on cl.id = c.claim_id
where cl.region = 'qld'
  and cl.status = 'open'
order by c.embedding <=> $1::halfvec
limit 5;

The Catch

Past about 10M vectors with high query concurrency you'll feel it. HNSW build times balloon, vacuum gets grumpy, and a specialised store like Qdrant will smoke you on raw QPS. If you're Spotify, you're not running pgvector. If you're literally anyone else, you probably should be.

The Verdict

Boring tech wins again. Pgvector + halfvec + Supabase is the single highest-ROI choice for any team that already has Postgres, which is every team. Ship the feature, save the infra migration for when you actually need it. In 2026, that bar is still very high.

Let us make some quick suggestions?
Please provide your full name.
Please provide your phone number.
Please provide a valid phone number.
Please provide your email address.
Please provide a valid email address.
Please provide your brand name or website.
Please provide your brand name or website.