Weaviate Review

The buying call is exit optionality. Vector databases became a budget line in 2024 and most teams grabbed Pinecone because it shipped first. Eighteen months later the same teams ask the board question: what happens if we want to leave? Weaviate''s answer is BSD-3 source on your own Kubernetes cluster.

Founded 2019 in Amsterdam by Bob van Luijt and Etienne Dilocker. Well-funded with NEA on the cap table. Hybrid Search and Multi-tenancy ship in the open-source core, not gated behind an enterprise SKU. The right shape for a category where lock-in is the loudest buyer concern.

The yellow flag is that Weaviate Cloud is younger than Pinecone''s managed offering and the support maturity gap shows. Pilot one production RAG workload for 60 days against your existing stack and keep the open-source escape hatch documented.

Three storage layers in one engine. Vectors in HNSW graphs, JSON objects alongside, BM25 inverted indexes in the same row — queryable in one GraphQL hop. That separates Weaviate from vector-only databases. Pinecone makes you bring your own keyword search; pgvector makes you wedge HNSW into Postgres.

The Modules system is the second strategic bet. text2vec-openai, generative-openai, reranker-cohere — each is a swappable pipeline stage. The catch is that Modules tie schemas to specific embedding providers, and migrating a billion-vector collection between models is a real backfill, not a config change.

Multi-tenancy at the schema level is the third choice that ages well. SaaS teams running per-customer RAG isolation get separate HNSW indexes per tenant, not row filters scanning the whole graph. Compare Qdrant where multi-tenancy is a 2024 retrofit, or Chroma where it''s effectively absent. Weaviate planned for this in 2021 and the design shows.

The TCO splits cleanly. Open-source: free, BSD-3, run it yourself. Weaviate Cloud Serverless: starts around $25/month for 1M vector dimensions. Enterprise Cloud: contact-sales — assume $50K-150K/year for a meaningful production cluster.

For a 50-person team running one RAG workload at 100M vectors, Weaviate Cloud lands $30-60K/year. Compare Pinecone Standard pods at roughly $70/month per p1 pod — math is similar, but Pinecone bills are easier to forecast because the unit is fixed pod count. Weaviate''s usage-based pricing rewards efficient schemas and punishes sloppy ones.

The hidden cost on self-hosted is the FTE. Budget 0.25-0.5 platform-engineer at $200K loaded, and your "free" deployment lands closer to $50-100K Year 1. The catch is that this is true of every self-hosted choice — Milvus, Qdrant, Vespa all carry the same human cost. Honest math against Pinecone, not cheaper than the GitHub README suggests.

Six PRs in and Weaviate stops feeling like a database and starts feeling like a framework. The Python v4 client is properly typed. Schema-first means you define your collection once and the client knows the rest. Hybrid Search is one ``near_text`` plus ``bm25`` call — compare Pinecone where keyword search is your problem to solve outside.

The hot path is the Modules system. text2vec-openai handles embedding at write time, generative-openai handles RAG completion at read time. Vector, prompt, and response flow through one query — fewer moving parts than Chroma plus LangChain plus a reranker. The catch is that switching text2vec-openai to text2vec-cohere on a production collection is a backfill, not a flip.

The hard fight is HNSW tuning. ef, efConstruction, maxConnections — a sprint to understand what those do to recall and latency. The docs don''t pretend defaults work past 10M vectors. Engineer-grade depth with engineer-grade cost.

Boot Weaviate with the Docker compose from the docs. Five minutes. Fine. Then you write your first schema and realize the schema is the product — class definitions, vectorizer modules, property configs. A lot of decisions before indexing a document. Chroma asks for none of that. Pinecone asks for none of that. Weaviate asks for all of it upfront, and you understand why three weeks later.

The polish is uneven. Weaviate Cloud console is functional, not delightful. CLI is sparse. The Python v4 client is a real improvement, but you''ll hit stack traces surfacing GraphQL errors you have to translate.

Three months in, the schema decisions are paying back. Hybrid Search just works. Generative Search lets you do RAG without a separate orchestration framework. Multi-tenancy holds up. The catch: nothing here is forgiving to people who skip the docs. If your team reads, right tool. If not, Chroma is friendlier.

The marketing is honest. Weaviate''s landing page leads with Hybrid Search, Multi-tenancy, and bring-your-own-vectorizer — three claims the product delivers. The 2019 founding and NEA-backed funding put the vendor past the risk window most 2023-vintage vector DBs still live in.

What worries me isn''t what they say. It''s what''s missing — an answer to whether vector databases stay standalone five years out. pgvector is eating the low end. Elasticsearch, MongoDB, and Snowflake all added vector search. "Good enough vector search inside the database we already run" wins 60% of the use cases bought in 2024.

Weaviate''s answer is deeper Hybrid Search, Generative Search, and Modules — features pgvector won''t replicate cleanly. The right bet, but the catch is the dedicated-vector-DB market is contracting while bundled vector search expands. Strongest open-source option in a category that ages poorly.