vector database

In depth

The principle

The principle rests on a single operation: turning a text into a series of numbers that situates its meaning, then comparing these positions with one another. Two passages that speak of the same thing in different words end up as neighbours, which no exact-word search allows. Querying the database therefore amounts to situating the question in the same way, then returning the closest passages. The database understands nothing: it measures proximities, and it is the model, further down the chain, that will read the passages returned.

What happens beforehand

The quality of a vector database is decided before the search, when the corpus is prepared. How the documents are cut up decides what can be retrieved: passages that are too short lose the context that made them intelligible, too long they drown the useful information in filler. The way texts are situated counts just as much, and changing it obliges you to redo everything, since the old positions are no longer comparable with the new ones. Finally, proximity of meaning ignores logic: a clause and its exception resemble each other, so do a procedure and its repealed version, and nothing in the measurement says which one is authoritative.

The trap

The first trap is to believe that resemblance amounts to relevance. A database always returns something, even when the answer exists nowhere in the corpus, and a model that receives an off-topic passage answers with the same assurance as if it were right. The second is forgetting permissions: indexing a corpus flattens documents which, in their system of origin, were not visible to everyone, and search by meaning does not restore those partitions by itself. The third is upkeep: a corpus that is fed without ever being pruned accumulates contradictory versions, and the system ends up faithfully quoting an obsolete text.