Vector databases overview

​

Supported databases

​

Feature comparison

​

Selection guide

​

Choose Pinecone if you need:

• Fully managed cloud service with zero infrastructure overhead
• Native sparse-dense hybrid search without external encoders
• Namespace isolation for 100,000+ tenants
• Serverless autoscaling for variable workloads

​

Choose Weaviate if you need:

• Native BM25 keyword search without sparse embeddings
• Generative search (RAG) with built-in LLM integration
• Multi-tenancy with per-tenant shards
• GraphQL query interface

​

Choose Milvus if you need:

• Partition-key isolation for millions of tenants
• JSON path indexing for complex metadata filtering
• Hybrid search with configurable RRF or weighted fusion
• Both Zilliz Cloud and self-hosted deployment options

​

Choose Qdrant if you need:

• Scalar or binary quantization for 4-32x memory reduction
• Maximal marginal relevance (MMR) for diverse results
• Tenant optimization for high-cardinality filtering
• High-performance gRPC protocol support

​

Choose Chroma if you need:

• Local persistent storage for development and testing
• Lightweight deployment without external dependencies
• Rapid prototyping with minimal configuration
• Tenant/database isolation for multi-tenancy

​

Multi-tenancy strategies

db.upsert(documents, namespace="tenant_1")
results = db.query(vector=embedding, namespace="tenant_1")

db.create_collection("Documents", enable_multi_tenancy=True)
db.create_tenants(["tenant_a", "tenant_b"])
db.with_tenant("tenant_a").upsert(documents)

db.create_collection("docs", use_partition_key=True, partition_key_field="tenant_id")
db.insert_documents(documents, namespace="tenant_1")
results = db.search(query_embedding=vec, scope="tenant_1")

db.create_namespace_index(namespace_field="tenant_id")
db.index_documents(documents, scope="tenant_1")
results = db.search(query_vector=vec, scope="tenant_1")

tenant_db = db.with_tenant("tenant_1", database="prod")
tenant_db.upsert(documents)

​

Hybrid search approaches

​

Pinecone and Weaviate: Native hybrid

# Pinecone: Native sparse-dense fusion
results = db.query_with_sparse(
    vector=dense_embedding,
    sparse_vector=sparse_embedding,
    top_k=10
)

# Weaviate: BM25 + vector with alpha balancing
results = db.hybrid_search(
    query="machine learning",
    vector=embedding,
    alpha=0.5  # 1.0 = vector only, 0.0 = BM25 only
)

​

Milvus and Qdrant: RRF fusion

# Milvus: RRF or weighted ranker
results = db.search(
    query_embedding=dense_vec,
    query_sparse_embedding=sparse_vec,
    ranker_type="rrf",  # or "weighted" with weights=[0.7, 0.3]
    top_k=10
)

# Qdrant: RRF via query fusion
results = db.search(
    query_vector={"dense": dense_vec, "sparse": sparse_vec},
    search_type="hybrid",
    top_k=10
)

​

Connection patterns

from vectordb.databases import PineconeVectorDB

db = PineconeVectorDB(
    api_key="pc-xxx",
    index_name="my-index"
)
db.create_index(dimension=768, metric="cosine")

​

Performance considerations

​

Memory optimization

• Qdrant offers the best memory efficiency with scalar (4x reduction) and binary (32x reduction) quantization
• Milvus supports large-scale deployments with partition-based data distribution
• Chroma is ideal for small to medium datasets in local environments

​

Query latency

• Pinecone and Qdrant provide the lowest latency with optimized indexing
• Weaviate excels at hybrid search with native BM25 integration
• Milvus offers configurable HNSW parameters for latency vs. recall tradeoffs

​

Throughput

• Qdrant uses gRPC by default for higher throughput than HTTP-based databases
• Milvus supports batch operations with configurable shard numbers
• Pinecone auto-scales for variable workload patterns

​

Next steps