Storing embeddings.
WHAT A VECTOR STORE DOES
Holds embeddings and finds the most similar to a query embedding.
WHAT SIMILARITY SEARCH IS
Finding nearest neighbours in the representation space.
WHAT EXACT SEARCH COSTS
Comparison against everything, which does not scale.
WHAT APPROXIMATE SEARCH PROVIDES
Near-nearest results, far faster.
WHAT IT TRADES
A small proportion of correct results, for large speed gains.
WHAT THE OPTIONS ARE
Extensions to existing databases Purpose-built vector databases Libraries embedded in an application
WHAT TO START WITH
An extension to a database you already run.
WHY
It avoids another system, and suffices at modest scale.
WHAT TO ESTABLISH BEFORE ADOPTING A SEPARATE SYSTEM
The actual corpus size and query rate.
WHAT FILTERING MATTERS FOR
Restricting results by attributes: permissions, date, source.
WHY IT IS A KEY SELECTION CRITERION
Filtering applied after retrieval returns too few results.
WHAT TO PREFER
Filtering applied during search.
WHAT HYBRID RETRIEVAL COMBINES
Keyword matching with similarity search.
WHY IT PERFORMS BETTER
Keyword search catches exact terms, identifiers and rare words that embeddings miss.
WHAT TO MEASURE
Whether the right passage appears in the results.
