Planned — Not Yet Implemented

This backend is on the roadmap but not yet implemented. The interface is designed for future additions. Currently only SQLite is available.

Qdrant

Purpose-built vector database with HNSW indexing, payload filtering, and distributed deployment.

Overview

Qdrant is a vector similarity search engine built from the ground up for vector workloads. It provides:

• HNSW index -- fast approximate nearest neighbor search at any scale
• Payload filtering -- filter vectors by metadata before or during search
• Distributed mode -- horizontal scaling with sharding and replication
• gRPC and REST API -- high-performance binary protocol and HTTP interface
• Qdrant Cloud -- fully managed service with free tier

When to Use

• Vector-first workloads where vector search is the primary access pattern
• Large-scale deployments with millions or billions of vectors
• Payload filtering -- complex metadata filters combined with vector search
• Managed cloud -- you want a hosted solution without managing infrastructure
• Multi-tenant -- collection-level isolation for different users or projects

Setup

Self-Hosted (Docker)

docker run -d --name qdrant \
  -p 6333:6333 \
  -p 6334:6334 \
  -v qdrant_storage:/qdrant/storage \
  qdrant/qdrant

Qdrant Cloud

Sign up at cloud.qdrant.io and create a cluster. The free tier includes 1GB of storage.

Create the Store

import "github.com/openparallax/openparallax/memory/qdrant"

store, err := qdrant.NewStore(qdrant.Options{
    URL:            "http://localhost:6333",
    CollectionName: "memories",
    Dimensions:     1536,
})
if err != nil {
    log.Fatal(err)
}
defer store.Close()

# Python
from openparallax_memory import QdrantStore

store = QdrantStore(
    url="http://localhost:6333",
    collection_name="memories",
    dimensions=1536,
)

// Node.js
import { QdrantStore } from '@openparallax/memory'

const store = new QdrantStore({
  url: 'http://localhost:6333',
  collectionName: 'memories',
  dimensions: 1536,
})

Qdrant Cloud Connection

store, _ := qdrant.NewStore(qdrant.Options{
    URL:            "https://your-cluster.qdrant.io:6333",
    CollectionName: "memories",
    Dimensions:     1536,
    APIKey:         os.Getenv("QDRANT_API_KEY"),
})

Collection Configuration

The store creates the collection automatically with sensible defaults:

store, _ := qdrant.NewStore(qdrant.Options{
    URL:            "http://localhost:6333",
    CollectionName: "memories",
    Dimensions:     1536,
    Distance:       "cosine",        // "cosine", "euclid", or "dot"
    OnDiskPayload:  false,           // store payloads on disk for large metadata
    HNSWConfig: &qdrant.HNSWConfig{
        M:              16,          // connections per node
        EfConstruct:    100,         // construction quality
        FullScanThreshold: 10000,    // switch to brute-force below this
    },
})

Payload Filtering

Qdrant supports filtering vectors by metadata (called "payloads") during search. Filters are applied before or during the vector search, not after -- this is efficient even with selective filters.

results, _ := store.Search(ctx, queryVector, 10,
    qdrant.WithFilter("source", "meeting-notes"),
    qdrant.WithFilter("date", qdrant.Range{GTE: "2026-01-01"}),
)

Supported filter types:

• Match -- exact value match
• Range -- numeric or date ranges
• Geo -- geographic bounding box or radius
• Has ID -- filter by point IDs
• Nested -- filter on nested payload fields
• Boolean -- AND, OR, NOT combinations

Full-Text Search

Qdrant supports full-text search on payload fields via text indexes:

store, _ := qdrant.NewStore(qdrant.Options{
    URL:            "http://localhost:6333",
    CollectionName: "memories",
    Dimensions:     1536,
    TextIndexFields: []string{"text"},  // create text index on "text" field
})

// Full-text search on payload
results, _ := store.SearchText(ctx, "CORS middleware", 10)

INFO

Qdrant's full-text search is basic compared to FTS5 or Elasticsearch. For advanced full-text needs, consider using Memory's hybrid search with application-level merging, or choose Weaviate for native hybrid search.

Performance

Benchmarks on a 4-core machine with SSD:

Vectors	Dimensions	Index Build	Query (top-10)	Memory
100,000	1536	45s	< 2ms	800 MB
1,000,000	1536	8 min	< 5ms	8 GB
10,000,000	1536	1.5 hrs	< 10ms	~80 GB

Qdrant supports on-disk storage with memory-mapped files for collections that exceed available RAM. Query latency increases slightly but remains practical.

Distributed Deployment

For large-scale deployments, Qdrant supports sharding and replication:

# qdrant-config.yaml
storage:
  performance:
    max_search_threads: 0  # auto-detect
  optimizers:
    default_segment_number: 4

cluster:
  enabled: true
  p2p:
    port: 6335

• Sharding -- distribute vectors across multiple nodes
• Replication -- replicate shards for fault tolerance
• Consensus -- Raft-based consensus for cluster coordination

Next Steps

• Choosing a Backend -- comparison with other backends
• Pinecone -- managed alternative
• PostgreSQL + pgvector -- ACID alternative
• Embeddings -- configuring embedding providers