Qdrant

Qdrant stores points made of one or more vectors plus a JSON payload, and filters on payload fields during the vector search rather than after it. Payload indexes marked as tenant keys keep each user's data together. For keyword matching, version 1.20 added BM25 scoring directly on a full text payload index, so one query can fuse BM25 and dense vector results without a separate sparse vector, plus a code tokenizer that matches identifiers such as getUserName and get_user_name. Sparse vectors (BM25 or learned models such as SPLADE) still work. The official mcp-server-qdrant package gives agents qdrant-store and qdrant-find tools.

As of October 28, 2026 the latest server release was v1.20.0 (October 22) and the Python client was qdrant-client 1.20.0 (Python 3.10 or newer), which also has an in process local mode for tests. Self hosted Qdrant has no authentication by default, so set an API key before exposing it; from 1.20, internal auth between cluster nodes is enforced by default once a key is set.

Pricing per qdrant.tech/pricing: the engine is free to self host. Qdrant Cloud has a free forever single node cluster (0.5 vCPU, 1 GB RAM, 4 GB disk) and usage based Standard tiers, plus Hybrid and Private Cloud options for your own infrastructure.