Speaker: Etienne, CTO and co-founder of Weaviate, a vector database (1:05:23).
The four (plus one) use cases
| Use case | What it is | Status |
|---|---|---|
| Semantic search | Search by meaning instead of exact keyword matching, via LLM-generated vector embeddings. Very common in e-commerce (1:06:02) | Core |
| RAG (retrieval-augmented generation) | The "ask AI" pattern — chat with your own data: Dropbox, Notion, email (1:06:02) | Core |
| Recommendations | Similarity search alone makes bad recommendations: "if you like this pepperoni pizza you'll like this pepperoni pizza" — similarity returns near-identical items, which isn't useful (1:06:02, 1:06:52) | Beta |
| Generative feedback loops | Apply generation on the write side, not just retrieval — a self-updating (or self-learning) database. Example: "translate everything to Spanish, summarize to 80 words, write it back" (1:06:52) | Alpha, working title |
Feature set (database-centric view)
Design premise stated by the speaker: architecture follows business functions, so features come first (1:07:36).
- Object and vector storage — Weaviate does not just return IDs to be looked up in a secondary system; objects live in the same database. Simplifies the end-to-end case, at the cost of its own challenges (
1:07:36,1:08:09). - Advanced filtering — equality, inequality, range (e.g. products in a price band), and minority filters (
1:08:09). - Hybrid search — combine keyword search with vector search, fused via e.g. reciprocal rank fusion (
1:08:42). - High updatability — trivial for a normal DB, genuinely hard for vector indexes. Real production case: 150 million vector updates per day in a single cluster (
1:08:42). - Distributed — replication, sharding, and offloading to cloud storage to decouple storage from compute (
1:08:42,1:09:15). - Performance focus — SIMD optimization, algorithmic optimization, and dedicated indexes per data type (
1:09:15).