Benchmarks

We wanted to be able to tell you just how good FEVER works. So we measured it!

Visual Embedding Throughput

  • CPU: ~50 images/sec No GPU quota required; use cheaper, available nodes for everyday usage.
  • GPU: ~2500 images/sec Go faster if you want to and can afford it! Make those backfills evaporate.
  • We measured these numbers on our production system deployed to AWS.
  • Query nodes run alongside indexing nodes - at its core we use Postgres for high scaleability, robust observability, and the security guarantees developers trust.

Media Embedding Throughput

  • Video: ~400 clips/sec on GPU, indexing and backfill included.
  • Documents / PDF: ~1,500 pages/sec on GPU - every page searchable, vector and full text.
  • Audio: ~50x realtime - an hour of audio indexed in about a minute.
  • Combined ingest: ~2,000 items/sec sustained.
  • Measured the same way as the visual embedding test against the same production system on AWS. Full hardware details and cost-per-corpus math are on the specifications page.

Seeing is Believing

  • If you want to verify our insane speeds, book a demo and we will show you directly how we can solve bigger problems than nearly anyone else, for less total cost.
  • And, if you are curious about end to end latency - it depends on your compute budget, instance types and sizing, and your database configuration. You control it all, we make it easy.

Try the example demo yourself

  • The live demo on this site runs against the same appliance we ship on AWS Marketplace.
  • We believe in, as the youth say "dogfooding," our products. Our demo reflects what you actually get - smart AI running on cheap commodity hardware!