A layer of Barrel

Barrel Vector

The vector layer. High-performance single-node vector search built in Erlang. Use it inside the full Barrel, or standalone as an embeddable library.

Choose Your Backend

Four indexing backends, one API. Pick the right one for your workload.

DEFAULT

HNSW

Pure Erlang implementation. Best for most use cases.

  • + Fast search <5ms
  • + No dependencies
  • + Good recall
GPU

FAISS

Meta's library via NIF. Best for large scale + GPU.

  • + GPU acceleration
  • + Billion-scale
  • + IVF, PQ compression
DISK

DiskANN

Pure Erlang SSD-optimized index. Best for large datasets on disk.

  • + SSD-optimized
  • + Low memory footprint
  • + No dependencies
KEYWORD

BM25

Pure Erlang full-text search. Combine with vectors for hybrid.

  • + Full-text search
  • + Hybrid with RRF
  • + No dependencies

Hybrid Search

Combine vector + BM25 with RRF fusion for best results

Multiple Metrics

Cosine, Euclidean, Dot Product distance

Metadata Filtering

Filter vectors by arbitrary JSON metadata

Persistence

RocksDB-backed storage with crash recovery

HTTP API

Simple REST interface for all operations

Embeddable

Use as a library in Erlang/Elixir apps

Performance

<5ms

Search (p99, top-10)

10K/sec

Insert throughput

1K QPS

Search throughput

Benchmarked on 1M vectors, 1536 dimensions

Quick Start

%% Start the vector store
{ok, _} = barrel_vectordb:start_link(#{name => docs, dimension => 3}).

%% Add a vector with metadata (or configure an embedder for text)
ok = barrel_vectordb:add(docs, <<"doc1">>, [0.1, 0.2, 0.3],
    #{title => <<"Hello">>}).

%% Vector search
{ok, Hits} = barrel_vectordb:search_vector(docs, [0.15, 0.25, 0.35], #{k => 10}).

%% Hybrid vector + BM25
{ok, Hits2} = barrel_vectordb:search_hybrid(docs, <<"hello">>, #{k => 10}).
# Reach the vector layer over HTTP by running barrel_server.
$ curl -X POST localhost:8080/db/docs/vector \
    -H 'content-type: application/json' \
    -d '{"id":"doc1","text":"Hello","vector":[0.1,0.2,0.3]}'

# Vector search.
$ curl -X POST localhost:8080/db/docs/search/vector \
    -H 'content-type: application/json' -d '{"vector":[0.15,0.25,0.35],"k":10}'

# Hybrid vector + BM25.
$ curl -X POST localhost:8080/db/docs/search/hybrid \
    -H 'content-type: application/json' -d '{"query":"hello","k":10}'

When to Use

Use Barrel Vector when:

  • + Single machine deployment
  • + <10M vectors
  • + Simple setup preferred
  • + Embedded in Erlang/Elixir app

Consider alternatives when:

  • + Need distributed cluster
  • + >10M vectors
  • + Automatic failover required
  • + Horizontal scaling needed
Read the Docs View Source