identity

Which model wrote these vectors, declared by you.

An entry records the identity of the embedder that produced it, and a lookup through a different identity refuses that entry instead of scoring it. See MissReason.EMBEDDER_MISMATCH. Two embedding models do not share a vector space, so a similarity computed across them is a number with no meaning, and a meaningless number near a threshold is exactly the condition that produces a false hit. Nothing throws and nothing appears in a log, because the dimension count usually matches: most upgrades within one provider's family keep it, which is why the existing dimension check catches only the loud version of this mistake.

kmemo does not try to infer this. There is no way to read a model's name out of its output, and a library that guessed would be wrong quietly. So you declare it, and what you declare is opaque to kmemo: it is compared for equality and never parsed. Include everything that changes the space: the provider, the model, the dimension count if the model is configurable, and your own revision marker if you fine-tune.

val embedder = object : Embedder {
    override val identity: String = "openai:text-embedding-3-small:1536"
    override suspend fun embed(text: String) = openAi.embed(text)
}

The default is UNDECLARED, and it is an identity like any other rather than a wildcard. A cache that declares nothing writes and reads undeclared entries and behaves exactly as it did before this existed. Declaring one on a store that already holds entries means those entries no longer match. That is the honest outcome, since nobody can say what wrote them, and docs/MIGRATION.md names the two ways through it.

Must be stable for the lifetime of the cache, for the same reason the vectors must be: it is read once, when the SemanticCache is constructed.