What to Expect: Approximate, Eventually Consistent, Not a General Database

Know the trade-offs

Vector databases make trade-offs you should know. Approximate: ANN search may miss true neighbours, so recall is a setting you tune. Consistency varies: some engines make a new write searchable only after a short delay or after an index segment is built (eventual consistency), while a transactional database like PostgreSQL gives normal ACID behaviour, including for vectors. Limited query power: dedicated engines are excellent at similarity plus filters but are not general-purpose analytics or transaction engines; joins and aggregations may be missing or basic. Scores are not probabilities: a similarity of 0.8 means different things for different models, so choose thresholds from your own data. Cost: memory dominates, since most fast indexes want vectors in RAM. Plan for these from the start rather than discovering them in production.

Test read-after-write on your engine

Insert a record, then query for it in a loop and measure how long until it appears. Do not assume zero delay.

Quick check: What does "eventual consistency" mean for a new write in some vector engines?

• It is searchable before it is written
• It may become searchable after a short delay
• It is never searchable
• It replaces the index