Vector Store Providers — 13 Backends

Beluga AI v2 provides a unified vectorstore.VectorStore interface for storing and searching document embeddings. All providers register via init() and are instantiated through the global registry.

The unified interface means your RAG pipeline works identically across all 13 vector store backends. Start with the in-memory store for development, switch to pgvector or ChromaDB for local testing, and deploy to Pinecone or Qdrant in production — all without changing your application code.

Interface

type VectorStore interface {
    Add(ctx context.Context, docs []schema.Document, embeddings [][]float32) error
    Search(ctx context.Context, query []float32, k int, opts ...SearchOption) ([]schema.Document, error)
    Delete(ctx context.Context, ids []string) error
}

Registry Usage

import (
    "github.com/lookatitude/beluga-ai/config"
    "github.com/lookatitude/beluga-ai/rag/vectorstore"

    // Register the provider you need via blank import
    _ "github.com/lookatitude/beluga-ai/rag/vectorstore/providers/inmemory"
)

func main() {
    store, err := vectorstore.New("inmemory", config.ProviderConfig{})
    if err != nil {
        log.Fatal(err)
    }

    err = store.Add(ctx, docs, embeddings)
    if err != nil {
        log.Fatal(err)
    }

    results, err := store.Search(ctx, queryVec, 10,
        vectorstore.WithThreshold(0.7),
    )
    if err != nil {
        log.Fatal(err)
    }
}

Available Providers

Provider	Registry Name	Type	Distance Strategies
pgvector	pgvector	PostgreSQL extension	Cosine, Dot Product, Euclidean
Pinecone	pinecone	Managed cloud	Cosine
Qdrant	qdrant	Self-hosted / cloud	Cosine, Dot Product, Euclidean
Milvus	milvus	Self-hosted / cloud	Cosine
Weaviate	weaviate	Self-hosted / cloud	Cosine
ChromaDB	chroma	Self-hosted	Cosine
Elasticsearch	elasticsearch	Self-hosted / cloud	Cosine
Redis	redis	Self-hosted / cloud	Cosine
MongoDB Atlas	mongodb	Managed cloud	Cosine
SQLite-vec	sqlitevec	Embedded (CGO)	Euclidean (L2)
Turbopuffer	turbopuffer	Managed cloud	Cosine, Dot Product, Euclidean
Vespa	vespa	Self-hosted / cloud	Cosine, Dot Product, Euclidean
In-Memory	inmemory	In-process	Cosine, Dot Product, Euclidean

Search Options

All providers support the same functional options for search configuration:

// Filter by metadata
results, err := store.Search(ctx, queryVec, 10,
    vectorstore.WithFilter(map[string]any{"category": "technical"}),
)

// Set minimum similarity threshold
results, err := store.Search(ctx, queryVec, 10,
    vectorstore.WithThreshold(0.8),
)

// Select distance strategy
results, err := store.Search(ctx, queryVec, 10,
    vectorstore.WithStrategy(vectorstore.DotProduct),
)

Provider Discovery

List all registered providers at runtime:

names := vectorstore.List()
// Returns sorted list: ["chroma", "elasticsearch", "inmemory", ...]

Hooks

Hooks allow observing vector store operations:

// BeforeAdd is called before each Add operation
// AfterSearch is called after each Search operation
hooks := vectorstore.Hooks{
    BeforeAdd: func(ctx context.Context, docs []schema.Document) error {
        log.Printf("Adding %d documents", len(docs))
        return nil
    },
    AfterSearch: func(ctx context.Context, results []schema.Document, err error) {
        log.Printf("Search returned %d results", len(results))
    },
}

Choosing a Provider

Use Case	Recommended Provider	Why
Testing and development	In-Memory	No setup required, deterministic
Existing PostgreSQL deployment	pgvector	Vectors co-located with relational data
Embedded/edge applications	SQLite-vec	File-based, no server needed
Managed cloud with zero ops	Pinecone, Turbopuffer	Automatic scaling, no infrastructure
Full-text + vector search	Elasticsearch, Weaviate	Combined keyword and semantic queries
High-performance self-hosted	Qdrant, Milvus	Purpose-built for vector workloads at scale
Existing Redis infrastructure	Redis	Low-latency, reuses existing deployment
Existing MongoDB Atlas	MongoDB Atlas	Vector search integrated with document DB
Production search engine	Vespa	Real-time indexing, custom ranking models