Weaviate

Weaviate's pitch is consolidation: instead of wiring a vector store to a separate embedding pipeline, a query layer, and an agent memory service, you get all four under one deployment. That matters in practice. The Query Agent turns natural-language questions into optimized database operations without you writing retrieval code, and since August 2026 its Search Mode adds medium, high, and ultrahigh effort tiers so you can trade latency for accuracy on hard queries. Engram, generally available since June 2026, gives agents memory that persists and adapts per user, which is the piece most teams end up hand-rolling badly. The database itself has kept pace. HFresh, a disk-based vector index aimed at billion-scale collections, reached GA in 1.38, and 1.39 promoted the Boost API and MMR diversity to GA while previewing 4-bit Rotational Quantization. Query profiling (July 2026) returns per-stage, per-shard timings, which is the kind of diagnostic surface vector databases usually lack. The MCP Server lets IDE assistants and LLMs interact with your instance directly — useful for AI-assisted coding and codebase search. Where the honest friction lies: the Free tier is a sandbox, not a workload. One collection, up to 3 tenants, 100,000 objects, 1 GB memory, 2,000 embedding requests/day, and 1,000 Query Agent requests/month. Flex at $45/mo removes the object limit and adds replication, but you still don't get SSO/SAML until Premium at $400/mo, and HIPAA compliance is Enterprise Cloud only. Pricing is metered by total vector dimensions, which is calculable but unusual — you'll want the Running Costs module in the console before committing. Data transfer is free only for a promotional period, so cost models should not assume it stays that way. Self-hosting via Docker or Kubernetes is genuinely open source and avoids all of this, at the price of running your own clusters and replication. Set against Pinecone or Qdrant, Weaviate gives you more out of the box — hybrid search, multi-tenancy at 50K+ tenants per cluster in customer reports, built-in embeddings — and asks for more operational attention in return. Choose it when search quality and tenant isolation are the product; skip it when you just need nearest-neighbor lookup on a few thousand rows.