Vector Database
A retrieval system optimized for storing vectors and searching nearest neighbors, usually alongside identifiers, metadata and filtering.
Retrieval Model
A vector database stores embeddings and retrieves items whose vectors are close under a configured metric. Exact search can be expensive at large scale, so systems often use approximate indexes such as HNSW.
Metadata filtering and vector search interact: filtering after ANN can reduce result quality, while pre-filtering can change index efficiency.
Operational Concerns
Index build time, memory footprint, update/delete behavior, recall, latency and persistence are part of the system design. A fast benchmark on one embedding distribution does not generalize automatically.
Embeddings should also be versioned when model changes alter the vector space.
Whether metadata filters run before, during, or after approximate search affects both latency and result quality. For mandatory authorization filters, removing unauthorized entries from the global top K does not guarantee that enough authorized candidates remain.
Related Information-Retrieval Concepts
- Embedding
- HNSW
- Cosine Similarity
- Approximate Nearest Neighbor
Related technical article: Artificial Intelligence: Philosophy, Theory and Practice.