A semantic search engine

Vela indexes your embeddings and answers queries by similarity in milliseconds, up to a billion vectors per index.

An index that is always current

You send vectors, Vela builds and maintains the index. k-nearest-neighbour search stays available while writes are happening.

The index is built on an HNSW graph spread across shards. Every write becomes queryable without manual reindexing. Namespaces isolate your datasets inside a single index.

12 ms p99 latency
1 B vectors per index
99.2 % recall

HNSW index

A neighbourhood graph built continuously

Cosine similarity

k-NN search up to 4096 dimensions

Data residency

8 regions to choose from for each index

Metadata filters

Queries combining vectors and attributes

Explore by function

The engine in detail, from vector ingestion through to the query.

Vector indexing

Embedding ingestion, HNSW graph construction and continuous updates

Ingestion

  • Batch upload, up to 4096 dimensions
  • Insert, update, delete
  • Writes queryable without reindexing
  • Namespaces per dataset

Index structure

  • Multi-layer HNSW graph
  • Automatic sharding
  • Up to 1 billion vectors per index
  • Background compaction

Embedding models

  • Works with any embedding provider
  • Dimensions declared at creation
  • Optional vector normalization
  • In-place reindexing when you change model

Similarity queries

k-nearest-neighbour search, scores and control over the recall/latency trade-off

k-NN search

  • Cosine similarity by default
  • Dot product and Euclidean distance
  • k set per query
  • Similarity score on every result

Recall control

  • Recall measured at 99.2 %
  • Adjustable exploration parameter
  • Recall/latency trade-off per query
  • Validation sets for comparison

Response format

  • Identifiers and scores
  • Metadata returned on request
  • Optional raw vectors
  • Cursor-based pagination

Metadata filters

Combining structured constraints and vector search in a single query

Field types

  • Strings and string lists
  • Integers and floats
  • Booleans
  • Timestamps

Operators

  • Equality and list membership
  • Numeric comparisons
  • AND / OR combinations
  • Negation

Execution

  • Filtering during graph traversal
  • Recall preserved on selective filters
  • Secondary indexes on frequent fields
  • Namespace as the first partition level

Performance

Measured latency, single-index capacity and scaling

Latency

  • p50 and p99 measured continuously
  • p99 at 12 ms
  • Measurements taken server side
  • Latency budget set per query

Capacity

  • 1 billion vectors per index
  • Up to 4096 dimensions
  • Shards added without downtime
  • Write throughput sustained during queries

Scaling

  • Read replicas per region
  • Automatic adjustment to query volume
  • Isolation per namespace
  • Quotas per API key

Operations

Regions, security and integration into your existing pipelines

Regions

  • 8 regions available
  • Data residency of your choosing
  • Optional cross-region replication
  • Multilingual search across 40 languages

Security

  • Encryption at rest and in transit
  • Scoped API keys
  • Exportable access logs
  • Private network access

Integration

  • HTTP and gRPC APIs
  • Python, TypeScript and Go libraries
  • RAG, product search, recommendation, deduplication
  • Documentation at docs.vela.dev

What we measure

Values recorded on our demonstration indexes.

12 ms

p99 latency

Response time for a k-NN search with a metadata filter, measured server side.

1 B

Vectors per index

Capacity of a single index, sharded automatically as writes come in.

99.2 %

Recall

Share of exact neighbours the HNSW graph finds on our validation sets.

40

Languages

Multilingual search over embeddings covering 40 languages in a single namespace.

Create your first index

Open an account, send your first vectors and run a similarity query. The documentation covers every call.