Weaviate

Architecture#

Weaviate runs as a standalone Go server with HTTP and gRPC endpoints. The default index type is HNSW per shard; flat indices are available for small collections. Storage uses a custom LSM tree built on top of RocksDB-like primitives, with separate stores for the object data and the vector index.

A Weaviate deployment is conceptually a list of 'collections' (formerly called 'classes'), each with typed properties and an optional vectoriser. Sharding is per-collection, with consistent hashing across nodes and raft-based metadata replication. As of 2026 the recommended production topology is at least three nodes for HA.

Modules#

Weaviate's most distinctive feature is its module system. Modules plug embedding models (text2vec-openai, text2vec-cohere, text2vec-transformers for self-hosted), rerankers (reranker-cohere, reranker-transformers), generators (generative-openai, generative-cohere, generative-anthropic), and other integrations directly into the database. The 'generative-*' modules execute RAG end-to-end inside Weaviate — query the collection, retrieve, and pass results to the configured LLM in one round-trip.

Hybrid Search#

Weaviate ships hybrid search as a first-class query parameter: `hybrid: { query: '...', alpha: 0.5 }` runs BM25 and dense in parallel and fuses with weighted score combination (alpha = 0 is pure BM25, alpha = 1 is pure dense). Reciprocal rank fusion is also available. BM25 is implemented natively with per-shard inverted indices.

import weaviate
from weaviate.classes.query import HybridFusion

client = weaviate.connect_to_local()
docs = client.collections.get("Documents")

result = docs.query.hybrid(
    query="how does HNSW handle deletions",
    alpha=0.5,
    fusion_type=HybridFusion.RELATIVE_SCORE,
    limit=10,
)
for obj in result.objects:
    print(obj.properties["title"], obj.metadata.score)

When to Pick Weaviate#

Weaviate is the most opinionated of the open-source vector databases. Pick it when you want a database that takes a position on schema, on hybrid search, and on direct LLM integration; when you are comfortable with Go-based ops; and when the module ecosystem aligns with your model providers. Pick something else when you want a lower-level, faster, less prescriptive vector engine — Qdrant tends to win those comparisons.