This PR merges the `vectors-optimize-brute-force` feature branch, which makes
the following changes to how vector functions are computed:
* Precompute the L2 norm of each vector at indexing time. (#45390)
* Switch to ByteBuffer for vector encoding. (#45936)
* Decode vectors and while computing the vector function. (#46103)
* Use an array instead of a List for the query vector. (#46155)
* Precompute the normalized query vector when using cosine similarity. (#46190)
Co-authored-by: Mayya Sharipova <mayya.sharipova@elastic.co>
Currently, when using script_score functions like cosineSimilarity, the query
vector is treated as an array of doubles. Since the stored document vectors use
floats, it seems like the least surprising behavior for the query vectors to
also be float arrays.
In addition to improving consistency, this change may help with some
optimizations we have been considering around vector dot product.
Currently when a document misses a vector value, vector function
returns 0 as a score for this document. We think this is incorrect
behaviour.
With this change, an error will be thrown if vector functions are
used with docs that are missing vector doc values.
Also VectorScriptDocValues is modified to allow size() function,
which can be used to check if a document has a value for the
vector field.
Typically, dense vectors of both documents and queries must have the same
number of dimensions. Different number of dimensions among documents
or query vector indicate an error. This PR enforces that all vectors
for the same field have the same number of dimensions. It also enforces
that query vectors have the same number of dimensions.
Julie Tibshirani
Mayya Sharipova