Weaviate

Weaviate is an open-source, cloud-native vector database developed by the Dutch company Weaviate B.V. (originally SeMI Technologies). First released in 2019, it has become one of the most widely adopted vector databases for retrieval-augmented generation (RAG), semantic search, and other AI applications that rely on high-dimensional embeddings. Weaviate distinguishes itself from pure vector search libraries by combining nearest-neighbour search with structured filtering, an integrated module system for automatic vectorisation, and a knowledge-graph data model that allows objects to reference one another.

Architecture

A Weaviate instance organises data into collections (formerly called classes). Each collection has a schema with typed properties — for example, a Document collection might have title, body, author, and createdAt fields. When an object is inserted, Weaviate stores both its scalar properties and its vector representation. Vectors can be supplied directly by the client or produced automatically through a configured vectoriser module that calls an embedding model such as OpenAI's text-embedding-3-small, Cohere's embed-multilingual-v3, or a self-hosted Hugging Face model.

For indexing, Weaviate primarily uses Hierarchical Navigable Small World (HNSW) graphs, with optional product quantisation, scalar quantisation, or binary quantisation to reduce memory footprint. A flat index option is available for small collections, and a dynamic index can switch between flat and HNSW as a collection grows.

Weaviate supports hybrid search that combines BM25 keyword scoring with vector similarity, fused via reciprocal rank fusion or a tunable alpha parameter. Filtering is first-class — queries can combine structured predicates (such as createdAt > 2025-01-01) with vector similarity in a single request.

APIs

Three protocols are exposed. REST is the original interface and is convenient for simple operations. gRPC, added in v1.23, is the recommended high-throughput interface for production workloads and is the default in the official Python client v4. GraphQL is offered as a query language particularly suited to cross-reference traversal between collections.

The Python, TypeScript, Java, and Go clients are officially supported, with several community-maintained clients in other languages.

Modules and ecosystem

Weaviate's module system handles vectorisation, reranking, generative answer construction, and other operations as pluggable components. Modules exist for OpenAI, Cohere, Hugging Face, Google Vertex AI, AWS Bedrock, Anthropic, Voyage, Jina, and several locally hosted alternatives such as text2vec-transformers and Ollama. Reranker modules (Cohere Rerank, Hugging Face cross-encoders) and generator modules (GPT, Claude, Gemini) allow Weaviate to act as an end-to-end RAG backend without external orchestration code.

Deployment

Weaviate can be self-hosted on Docker Compose, Kubernetes (with an official Helm chart and operator), or bare metal. The company also operates Weaviate Cloud, a managed multi-region service, and offers a serverless tier and a bring-your-own-cloud option for customers with data residency requirements. Production deployments rely on sharding for horizontal scaling, replication for high availability, multi-tenancy for SaaS architectures, and role-based access control (RBAC) for security.

Comparison with alternatives

| Database | Type | Strengths | |---|---|---| | Weaviate | Open source, managed | Hybrid search, modules, multi-tenancy | | Pinecone | Managed only | Serverless scaling, simple API | | Qdrant | Open source, managed | Strong filtering, payload-aware indexing | | Milvus | Open source, managed | Large-scale GPU acceleration | | Chroma | Open source, embedded | Developer ergonomics, lightweight | | pgvector | PostgreSQL extension | Familiar SQL, transactional |

Weaviate is commonly chosen when teams want a fully open-source option with first-class hybrid search, modular integration with multiple embedding providers, and a managed-cloud upgrade path.

Use cases

Typical Weaviate deployments include retrieval-augmented generation backends for enterprise chatbots, semantic search across document corpora, product and content recommendation, image and multimodal search, and knowledge graph-style applications that combine vector similarity with structured relationships.

Limitations

Vector databases including Weaviate carry operational complexity that should not be underestimated. Schema design, embedding choice, chunking strategy, and reranker selection have outsized effects on retrieval quality and are not solved by the database itself. Memory cost for very large collections can be significant unless quantisation is configured. As with any open-source platform, self-hosting requires capacity for upgrades, monitoring, and security patching that some teams prefer to outsource to managed services.

References

1. Weaviate B.V. (2025). Weaviate Documentation. docs.weaviate.io.
2. Malkov, Y. A., & Yashunin, D. A. (2018). Efficient and Robust Approximate Nearest Neighbor Search Using HNSW Graphs. IEEE TPAMI.
3. Weaviate B.V. (2024). Weaviate 1.27 Release Notes. weaviate.io/blog.
4. Zilliz. (2025). Top Open Source Vector Databases in 2025. zilliz.com.