Weaviate

💰 Pricing & Plans

🔧 API & Developer Experience

• •Official clients in five languages: weaviate-client for Python (v4.24.1+, needs Weaviate 1.27+), TypeScript/JS (v3.14.0+), Java (client6 v6.3.1+), C# (v1.2.0+) and Go (v5) — all talking gRPC under the hood rather than REST.
• •gRPC data path, REST metadata: Data operations go over gRPC on port 50051, metadata and management over REST on 8080. On Weaviate Cloud a single REST Endpoint URL covers both, so you never configure ports yourself.
• •GraphQL is deprecated: The GraphQL API is deprecated and is switched off on Weaviate Cloud on May 1, 2022027; clusters idle on GraphQL for 30 days lose it on December 1, 2026. New clusters ship with DISABLE_GRAPHQL=true, so plan on gRPC clients or the REST search endpoints.
• •Built-in MCP server: Set MCP_SERVER_ENABLED=true and the server exposes weaviate-collections-get-config, weaviate-tenants-list, weaviate-query-hybrid and weaviate-objects-upsert at /v1/mcp on the REST port. It is always on for Cloud clusters and can be locked to read-only tools.
• •Auto-schema and named vectors: Cloud enables auto-schema by default, so you can ingest before defining properties. Configure.Named vectors let one collection carry separate title and body representations and query each with target_vector.
• •RBAC with per-collection roles: RBAC is on by default from Weaviate v1.30. Create a role scoped to one collection with Permissions.collections and Permissions.data, then assign it to a user via client.users.db.assign_roles — no Admin key required in production.
• •Built-in agentic layer: The Query Agent turns natural language into Weaviate operations via the weaviate-agents package (v1.8.0+), handling retrieval, reranking and generation; Engram adds persistent semantic memory for agents.

🗄️ One Database, Hybrid Search & Built-in Embeddings

Weaviate is not a vector sidecar bolted onto an existing stack. It stores objects, vectors and the inverted index inside one LSM-backed database, so structured filters, sorting and BM25 keyword search run against the same records the vector index serves. That is why hybrid search is a single call - col.query.hybrid(query=..., limit=3) fuses lexical and vector scores with no second datastore to reconcile. Weaviate Embeddings closes the loop: the cluster that stores the vectors also generates them, so a RAG deployment needs no third-party API key for vectorization. Compression (RQ-8, PQ, SQ, BQ) trades a little recall for large memory and storage savings, and named vectors let one collection carry separate title and body representations.

🌐 China Access & Latency

Weaviate Cloud has no mainland-China region. Free clusters run on AWS only, Flex and Premium Shared cover a limited set of AWS and GCP regions, and only the Premium Dedicated tier reaches roughly 40 regions across AWS, GCP and Azure - so a China workload must pick Hong Kong, Tokyo or Singapore and accept cross-border latency on the gRPC data path. Measured from an Alibaba Cloud Malaysia host, weaviate.io and docs.weaviate.io answer in roughly 30-50ms via Netlify's AWS edge, while console.weaviate.cloud on GCP takes about 680-740ms. For production inside mainland China, self-hosting is the better answer: the server is BSD-3-Clause outside the wl/ directory, so you can run Weaviate on an Alibaba Cloud or Tencent region for near-domestic latency - but Weaviate Embeddings and the Query Agent are Cloud-only and must be replaced with a local vectorizer such as text2vec-ollama.