VectorSimilarity.INNER_PRODUCT requires that vectors be normalized in order to produce satisfying results.

We currently (as of ) warn about this in the documentation, but wouldn’t it be better to perform the normalization automatically?

• We’d expose @Vector(normalization = ...) (type float?), defaulting to -1 (meaning “default for the chosen similarity function”)

• For VectorSimilarity.INNER_PRODUCT, we’d normalize to 1.0f by default for float vectors, and to some other

• For other similarities, normalization would be disabled by default.

• Users could set @Vector(normalization = 0.0f) to disable normalization (if they already normalize their data), or @Vector(normalization = 232f) to set an arbitrary norm.

• Setting explicitly a norm that is not 0.0f or 1.0f with VectorSimilarity.INNER_PRODUCT and float vectors would result in an exception on startup: if people want to do something risky, they must disable our normalization and do it themselves.

Alternatively, we could only expose @Vector(normalize = VectorNormalize.YES/NO/DEFAULT) and pick a hardcoded norm for both types: 1.0f for floats, and maybe the maximum possible norm for bytes? I’m not sure how useful it is to set an arbitrary norm.