Qdrant Cloud

Qdrant Cloud gives you managed Qdrant without running a server — create a cluster, get an API endpoint, and use the same Qdrant Python/JavaScript client you would use against a self-hosted instance. Qdrant is known for high performance (Rust-based), advanced filtering (payload filters applied during vector search), sparse+dense hybrid search, and a clean API. For agents: create collections, upsert points (vector + payload metadata), and search with semantic similarity + metadata filters. Generous free tier (1GB, 1 node). Confidence is docs-derived.

Python client: pip install qdrant-client. Connect: from qdrant_client import QdrantClient; client = QdrantClient(url="https://xxx.qdrant.io:6333", api_key="api-key-here"). Create collection: client.create_collection(collection_name="docs", vectors_config=VectorParams(size=1536, distance=Distance.COSINE)). Upsert: client.upsert(collection_name="docs", points=[PointStruct(id=1, vector=[0.1, 0.2, ...], payload={"text": "...", "source": "web", "date": "2025-01-01"})]). Search: results = client.search(collection_name="docs", query_vector=[0.1, 0.2, ...], query_filter=Filter(must=[FieldCondition(key="source", match=MatchValue(value="web"))]), limit=5). Response: [{id, score, payload}]. Batch: client.upload_points(collection_name="docs", points=points_list, batch_size=256). Named vectors (hybrid): VectorsConfig({"dense": VectorParams(...), "sparse": SparseVectorParams(...)}). REST API: POST /collections/{name}/points, POST /collections/{name}/points/search.

API key auth: api-key header (REST) or api_key parameter (Python client). Keys from cloud.qdrant.io → Cluster → API Keys. HTTPS enforced. gRPC: also available on port 6334 with TLS. Namespace isolation: collections act as namespaces; no cross-collection queries. Multiple API keys: create per-service read/write keys. No OAuth2. Qdrant Cloud: SOC 2 Type II certified. Self-hosted Qdrant: configurable API key or no auth (internal only).

Search latency: ~5–20ms for typical collections (1M vectors, 1536 dims). Indexing: HNSW index built automatically; large collections may take minutes to fully index. Upsert: batch upserts (batch_size=256) significantly faster than single-point inserts. Filtering: payload filters applied during HNSW search (no post-filter performance penalty). Quantization: scalar/binary quantization reduces memory usage 4–32x with minor recall loss. Collection size: free tier 1GB; paid tiers scale to TB. Backup: automatic daily snapshots on Qdrant Cloud. Payload indexing: index payload fields for faster filtering (client.create_payload_index(...)).

qdrant.tech/documentation covers Python/JavaScript/Rust client, collection management, search options, filtering, quantization, and cloud setup. Getting started: create cluster at cloud.qdrant.io (free tier) — first vector search in 5 minutes. Free tier: 1GB, 1 node (sufficient for ~1M 1536-dim vectors). GitHub: qdrant/qdrant (22k+ stars). Community via Discord (very active). Exceptional documentation; Qdrant has the most thorough vector search optimization guide available.