Managed vector search and retrieval for engineering teams. 24 ms p99 at a billion vectors, recall you can measure, and no index for your team to babysit.
Three things go wrong with a self-run vector index, and they all go wrong at the same size.
p99 triples the week you cross fifty million vectors, and nobody on the team owns the index.
Approximate search quietly drops the right answer. You hear about it from a user, not a dashboard.
Shards, rebuilds, replica lag, memory ceilings. Two engineers end up running a store nobody set out to build.
One retrieval layer with a published latency number, a recall report and nothing for you to operate.
We pick the index family, shard count and replica layout, then keep them inside your latency budget.
Dense vectors, keyword BM25 and metadata filters resolved together, in a single round trip.
Single-tenant indexes, encryption at rest and in transit, and a region you choose.
Point us at Postgres, S3 or a queue. New rows are embedded, indexed and searchable in seconds.
24ms
p99 Query Latency
99%
Recall @10, Scale Tier
12B+
Vectors Indexed
40K/s
Peak Queries Served
[ how it works ]
From zero to production in three steps
Step 1
Push vectors over REST or gRPC, or point us at Postgres, S3 or Kafka. We build the index.
Bulk import at up to two million vectors a minute
Schema and dimension checks before anything is indexed
First query answered inside five minutes
Step 2
One endpoint for approximate search, metadata filters and keyword hybrid, with clients for Python, TypeScript and Go.
Per-query latency, recall and cost in the response headers
Filters applied inside the index, not after it
Read replicas in the regions your users are in
Step 3
Give us a recall target and a latency budget. We tune the index against your own traffic and show the working.
Weekly recall report against exact ground truth
Automatic rebuilds with no read downtime
A named engineer on the Scale tier and above
[ Testimonials ]
Trusted by teams shipping AI products at scale.
"We moved 300 million embeddings onto Nodeform over a weekend and our p99 dropped from 180 ms to 22 ms. Nobody has touched an index since."
Nadia Farouk
Staff Engineer, Arclight Data
"The recall report is the part I did not know I needed. We can finally tell product why an answer was missed."
Callum Reid
VP Engineering, Cobalt Works
"We went from 40 million vectors to 1.4 billion in nine months. No rebuild, no migration weekend, no config change."
Wei Zhang
ML Platform Lead, Tessell Labs
"Hybrid filters in the same query killed a whole service for us. One endpoint now does what three of ours used to do badly."
Joana Ribeiro
Head of Search, Kestrel AI
[ blogs ]
Benchmarks, postmortems and notes on running vector search in production.
View all posts
