cluster healthy · p95 47ms
cluster · 1.4b vectors · 6 regions

Embeddingsat billion-scale,answered in 47ms.

Spectra is a vector database for serious retrieval. Topical sharding, native hybrid search, in-process reranker, and a per-tenant model that does not pretend isolation is a metadata filter.

Spin up indexView SDK
47ms
p95 query
1.4B
vectors
94.6%
recall@20
query · semantic / k=20
47ms · 1.4B vec
"a wooden lighthouse first lit in the 1860s on the north sea"
0.94
Mariners Library / Bk II / p.47
Wooden lighthouse, north pier, first lit 17 March 1864…
0.91
Coastal Survey 1871 / sheet 4
Pier light marked at 53.4N, replaced 1882 with stone tower…
0.86
Trustees Letter / 1864-04-12
Lamp procured from Edinburgh foundry, first kindled by night-watch…
0.81
Editorial 1907 / p.213
Old wooden post still recalled by elderly residents of harbour…
0.78
Photo plate 1899 / album C
Showing pier and timber tower, possibly the original 1864 fixture…
47ms
p95
20
k results
+rerank
0.94 NDCG
Northwind Labs✦Mercury OS✦Substrate✦Foundry 47✦Plover Studio✦Helvetia Cartographic✦Field Notes Co.✦Linnea Bakery✦Marlow Workwear✦Verge Editions✦Northwind Labs✦Mercury OS✦Substrate✦Foundry 47✦Plover Studio✦Helvetia Cartographic✦Field Notes Co.✦Linnea Bakery✦Marlow Workwear✦Verge Editions✦
02 / Cluster

Four organs. One nervous system.

Topical shards

Vectors that get queried together live together. p95 stays flat as the corpus grows.

Hybrid retrieval
sparse · BM25
"lighthouse" "1864" "wooden"
dense · 1024d
[0.32, -0.18, 0.41, ... ]
→ fused @ rank, k=200
Reranker
+8.3% NDCG
vs unranked
●in-process
●8.4ms p95
●finetune-able
Multi-tenant
tenant-a
412M vectors
isolated
tenant-b
184M vectors
isolated
tenant-c
47M vectors
isolated
tenant-d
12M vectors
isolated
tenant-e
8M vectors
isolated
tenant-f
1.4M vectors
isolated
Geography
us-east-1
22ms
us-west-2
47ms
eu-west-1
38ms
eu-central-1
41ms
ap-southeast-1
184ms
ap-northeast-1
196ms
03 / In production

Three workloads. One cluster.

Search, RAG, personalisation. Same query, same SLA, regardless of which one you reach for.

01
Search

Semantic search over a billion vectors. p95 under 47ms.

Spectra shards by topic, not by hash. Hot vectors stay in memory; the long tail spills to NVMe with predictable tail latency. Same API, any scale.

0
ms p95 across 1.4B vectors
02
RAG infra

Retrieval that actually answers the question.

Hybrid sparse + dense retrieval, learned reranker, and a citation surface that points to the source paragraph. Plug it into any model gateway in eight lines.

0
% recall @ 20 on long-tail
03
Personalisation

Per-user vector spaces without per-user infrastructure.

Multi-tenant indexes with strong isolation. Onboard a customer in 8.3 seconds. Tear down in 8.3. No noisy-neighbor surprises during your launch week.

0
s tenant onboard / teardown
04 / Vs scaffolded vector DB

Built for production. Not for demos.

Capability
Spectra
Generic vector DB
Latency at billion-scale
47ms p95
240ms p95
Hybrid retrieval
Sparse + dense, native
Bolt-on
Reranker
Learned, served in-process
Separate service
Tenant isolation
Per-tenant index
Shared with metadata filter
Replication
Multi-region, write-anywhere
Single primary
field reports

“Spectra ate our 1.4 billion-vector corpus and answered in 47 ms p95. Our previous setup needed a Kubernetes cluster and a prayer.”

Soraya Mehta
Search Lead, Northwind Labs
0B
vectors served
0ms
p95 query
0%
recall @ 20
0
active regions
05 / Pricing

Pay for vectors. Not for nodes.

Bench
$0free for hacks

For prototypes and learning.

  • 1M vectors
  • 1 index
  • Read-replica only
  • Community Discord
Spin up an index
Lab
$184per project / month

For teams shipping retrieval features.

  • 120M vectors
  • Multi-tenant indexes
  • Hybrid + reranker
  • Two regions
  • API + webhooks
  • Priority email
Start 14-day trial
Foundry
—annual contract

For billion-vector and regulated workloads.

  • Unlimited vectors
  • Dedicated cluster
  • VPC peering, BYOK
  • Multi-region active-active
  • Solutions architect
  • 99.96% SLA
Talk to founders
06 / FAQ

Engineering answers to engineering questions.

We shard topically — vectors that get queried together live together — and keep hot shards memory-resident. The long tail spills to NVMe with a predictable miss penalty. p95 stays flat as the corpus grows.

Recent shipments

all releases
Apr 26
index
Topical sharding now auto-rebalances when query patterns shift > 12%.
Apr 17
rerank
Native reranker upgraded — 8.3% recall improvement on long-tail queries.
Apr 06
regions
ap-northeast-1 active-active replication generally available.
Mar 24
sdk
Python SDK 4.7 ships streaming deletes and bulk upsert chunking.

Index a million vectors on the house.

One CLI command. One region. Zero credit card. Scale when you have customers.

Spin up free Talk to engineer