VectorStores

Vector databases (Qdrant and others): indexing, similarity search, hybrid (vector + keyword) search, and metadata filtering.

Why we need this / what value this brings

Similarity search over embeddings is what makes RAG and semantic search fast at any real scale — a linear scan doesn't hold up past a few hundred documents.

When to use this

As soon as RAG or 'find similar X' functionality is needed over more than a trivial dataset.

How to use or implement this

Pick a vector DB (Qdrant, pgvector for smaller scale), choose an index type (HNSW is the common default), and decide what metadata filters (category, date) need to accompany the vector search.

Research questions

• HNSW vs IVF indexing — recall/speed/memory trade-offs at Localz's expected scale.
• Hybrid search: when does keyword search outperform pure vector similarity?
• How do you keep an index fresh as source documents change (re-embedding strategy)?

---

Empty folder — drop notes, links, and findings here as you research.