A document batch over one of the engine’s batch limits (items, chunks, source or metadata
bytes, embedding values, transaction bytes). Nothing was written; a smaller batch can succeed.
When a document’s chunk hypervectors are computed. Terms, postings and text are stored at
ingest in every mode, so lexical search sees the document at once; only the vector side
(HmsCore::query over chunk vectors) differs.
Default beam width of a graph query. The accepted qgraph rows reached recall 0.95 at
ef 736 on glove-100-angular (1.18M vectors, benchmarks/results/ann_modal_confirm_v1.json);
stores of up to a few hundred thousand embedding rows reach it with a far smaller beam, and
tests/server_ann.rs holds the default to recall >= 0.95 on its corpus.