You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Problem. Vectors are Value::Vector(Arc<Vec<f32>>) (src/value/mod.rs:222) and distance/LSH functions are pure (src/vector_ops.rs), but no maintained view exists to carry them, there is no centroid aggregate, and nothing proves that vector-carrying views stay exact under embedding updates.
Target. Views carry vector columns through arrangements by shared pointer (one copy per graph per distinct vector); pure vector functions (cosine, dot, euclidean, int8 variants, normalize, scale/add, lsh_bucket, lsh_probes) allowed in deployed rules and compiled into the maintainer like any function; a vec_mean aggregate (centroid of a group) maintained by reduce; an embedding update is a row retraction plus insertion through every view.
Acceptance. Oracle on random histories with embedding updates: vector-carrying views equal recompute at every revision; memory test: two views carrying 100k vectors add <= 32 B per row over the base store; vec_mean retraction correctness.
Activity