NEWS

bigANNOY 0.3.0 (2026-04-01)

Add a true loaded-index lifecycle for the native C++ backend so repeated searches can reuse a live Annoy handle instead of reloading from disk each call.
Add annoy_is_loaded(), annoy_close_index(), and annoy_validate_index() for explicit operational control over persisted indexes.
Extend index metadata with an index id, file size, file modification time, file checksum, and load mode, and validate persisted files against those recorded properties.
Support eager and lazy load modes for built and reopened indexes.
Accept bigmemory descriptor objects and descriptor file paths anywhere the package accepts big.matrix references, queries, or streamed output destinations.
Expand the benchmark layer to support user-supplied datasets, self-search versus external-query benchmarks, saved CSV summaries, and validation of built indexes as part of the benchmark workflow.
Add tests covering eager/lazy loading, explicit close/reload, descriptor inputs, corrupted metadata, and the richer benchmark interface.

Make the native C++ bigmemory backend the default execution path and keep the R implementation as a debug-only fallback.
Add explicit annoy_open_index() and annoy_load_bigmatrix() helpers so build and query workflows can be decoupled across sessions.
Write and read sidecar metadata files alongside Annoy indexes, recording metric, dimension, item count, tree count, package version, Annoy version, build seed, build threads, and build backend.
Extend metric support beyond Euclidean to include "angular", "manhattan", and "dot".
Add v2 build/search controls including seed, build_threads, prefault, and block-size handling.
Strengthen file-backed and persistence workflows for reference matrices, destination matrices, and reopened indexes.
Add exported benchmark helpers, benchmark_annoy_bigmatrix() and benchmark_annoy_recall_suite(), plus an installed benchmark runner in inst/benchmarks/.
Expand test coverage for file-backed matrices, separated-column matrices, streaming outputs, metadata reload behavior, and native-versus-debug backend parity.

First public version of the package for approximate nearest-neighbour search on bigmemory::big.matrix data with Annoy.
Add the initial public API with annoy_build_bigmatrix() and annoy_search_bigmatrix().
Support exact package-level result shapes aligned with bigKNN, including index, distance, k, metric, n_ref, n_query, exact, and backend.
Support Euclidean Annoy search for self-search and external-query workflows.
Accept reference inputs as big.matrix objects or external pointers, and query inputs as NULL, dense numeric matrices, big.matrix objects, or external pointers.
Add streamed output support through destination big.matrix objects for neighbour indices and distances.
Introduce the first native C++ bigmemory path for block reads and Annoy build/search loops, alongside an R fallback backend.
Add initial correctness, validation, and backend-parity tests plus an installed benchmark script for moderate-size comparisons.