Approximate search silently drops the result that mattered.
Coherence is GPU-native exact k-nearest-neighbor search — recall = 1.0, in-memory, no index. The true nearest neighbors, not an approximation. The alternative to approximate ANN (HNSW, IVF, ScaNN).
Exact top-10 over 24M vectors @ 2,560-d, recall 1.0, no index, ~50 ms/query on one B300 — bit-exact vs a CPU oracle.
Use Coherence when approximate isn't good enough
You run retrieval, RAG, or agents in production — and one of these is your problem.
"A missed result is a liability."
Legal, compliance, clinical, financial, fraud, security. ~95% ANN recall is a non-starter when the 5% you miss is the document that mattered. Coherence returns the true nearest neighbors — recall 1.0.
"Your index lags every time the corpus changes."
Every re-index is a window where a new vector isn't searchable yet — the fraud "blind spot" an adversary lives in. Coherence has no index, so a new vector is searchable the instant it lands.
"You're paying to build, warm, and rebuild an index."
ANN means an index you build, hold in memory beside the vectors, tune, and rebuild on every update. Coherence has no index at all — nothing to build, warm, tune, or drift.
"You need to run and audit retrieval yourself."
Coherence is source-available. Run it on your own hardware, air-gapped, and read every line. Auditable source you run in your own air-gap is a stronger guarantee than a sealed binary.
Exact is back
"Approximate search was a workaround for hardware that couldn't do exact fast enough. The hardware caught up. The workaround didn't have to."
You'd never run a database that rebuilds its index on every write and hands back results you can't reproduce — anywhere else in your stack, that's dead on arrival. Approximate search asks you to accept exactly that: an index to rebuild on every update, recall below 1.0, and matches it quietly misses. It was the only option when exact search was too slow — and that stopped being true. Coherence returns the true nearest neighbors on every query, reproducibly, with no index to rebuild. Forge embeddings produce the vectors; Coherence finds them.
The benchmark
Measured, not projected. Bit-exact against a CPU brute-force oracle.
Single-query latency, measured on one GPU and verified bit-exact against a CPU oracle. Batching amortizes the corpus read across many queries.
Exact vs Approximate
An honest comparison against approximate indexes like HNSW, IVF, and ScaNN.
| Coherence (exact k-NN) | ANN (approximate) | |
|---|---|---|
| Recall | 1.0 — the true nearest neighbors | < 1.0 — silently misses results |
| Index | None — nothing to build or tune | Build, warm, hold in memory, rebuild |
| New vector | Searchable the instant it lands | Rebuild window before it's searchable |
| Cost model | Pay for the search, not an index | Index build + memory + drift overhead |
| Trade | Corpus must fit resident in GPU memory | Accuracy traded away to dodge the scan |
Source-available. Run it yourself.
Knowledge is precious — so the crown-jewel infrastructure is yours to read, audit, and run.
Read & audit
Source-available under the Functional Source License (FSL-1.1-Apache-2.0). Read it, audit it, modify it.
Self-host & air-gap
Run it on your own hardware, including fully air-gapped, free, for any purpose except reselling it as a competing hosted service.
Converts to Apache 2.0
Two years after each release, the code converts to Apache 2.0. Defended now, fully open later.
Regulated buyers want to audit and run crown-jewel infrastructure themselves. Open, auditable source you run in your own air-gap is a stronger guarantee than a sealed binary.
Become a Design Partner
The exact retrieval core is built and measured. We're accepting Design Partners for early access while we productize.
NVIDIA Inception program member | Seed Stage