Weaviate Vector DB

Open-source vector database providing semantic and hybrid search infrastructure. Supports native multi-stage retrieval pipelines including ANN retrieval and cross-encoder reranking. Maintained by Weaviate (the company).

• Website: https://weaviate.io/
• GitHub: https://github.com/weaviate/weaviate
• Cloud: Weaviate Cloud (WCD) — managed hosted service

---

What It Does

Weaviate stores objects with vector representations and exposes search via GraphQL and REST APIs. It integrates vectorization and reranking as first-class pipeline stages rather than requiring external tooling.

Key capabilities:

• ANN search — HNSW-based index
• Hybrid search — combines dense and sparse (BM25) retrieval with fusion
• Native reranking — cross-encoder reranking built into the search pipeline (no separate service)
• Vectorizer modules — pluggable embedding providers (OpenAI, Cohere, HuggingFace, etc.) called at import and query time
• Multi-modal — image + text vectors in the same collection
• GraphQL API — declarative query syntax for retrieval, filtering, aggregation

Two-Stage Pipeline

Weaviate’s canonical retrieval pattern:

1. Stage 1 — Bi-encoder ANN: fast candidate retrieval using pre-computed embeddings (Bi-Encoder)
2. Stage 2 — Cross-encoder reranking: accurate reranking of top-K candidates using query-aware scoring (Cross-Encoder)

This separates scalability (stage 1) from accuracy (stage 2).

Multi-Vector Embeddings and MUVERA

Since 1.31, Weaviate can encode multi-vector embeddings (ColBERT, ColPali) with MUVERA so HNSW indexes one fixed-size vector per object, tuned with k_sim, d_proj and r_reps (no final projection). On LoTTE lifestyle with colbertv2.0, memory dropped from ~12 GB to under 1 GB and import from 20+ minutes to 3–6 minutes, at the cost of recall that needs a higher ef (above 512 for 80%+) to recover. See More Efficient Multi-Vector Embeddings with MUVERA.

Related Tools

• Qdrant Vector DB — competing vector database; stronger quantization options (TurboQuant)
• Pinecone — managed-only competitor
• FAISS — library only; no pipeline features

Related Concepts

• HNSW — ANN index used
• Hybrid Search — dense + sparse fusion
• Cross-Encoder — reranker Weaviate integrates natively
• Bi-Encoder — first-stage retriever
• Reranking — the two-stage pattern Weaviate popularized
• Dense Vector Retrieval — primary retrieval mode

Articles

• Using Cross-Encoders as Reranker in Multistage Vector Search

• Choosing a Vector Database for ANN Search at Reddit — Reddit qualitatively scored Weaviate; not tested quantitatively due to time constraints; single-node-type architecture noted as a concern

• More Efficient Multi-Vector Embeddings with MUVERA — MUVERA support in 1.31, LoTTE memory/import/recall benchmark

People

• Laura Ham — cross-encoder pipeline explainer

• Roberto Esposito · Joon-Pil (JP) Hwang — MUVERA implementation post