USearch by Unum Cloud

Release: v2.25.0 [skip ci]

Minor

• Add: Native UInt8 u8 API and Float6 e2m3 & e3m2 quantization (e4db172)
• Add: Float8 support with E5M2 & E4M3 (bf80af8)
• Add: Uniform hardware caps queries across SDKs (11f4d5c)
• Add: expose memory_stats through Rust FFI bindings (d5a04f8)
• Add: memory_stats_t struct with per-tape breakdowns (8f6f8f6)

Patch

• Fix: Harden min & max symbols for MSVC (a270651)
• Fix: Over-allocate & in-align striped locks (7aee20f)
• Improve: Print stack-traces in test singal-handlers (94c7ec9)
• Make: Bump NumKong to v7.5 (c666f16)
• Docs: Benchmarks table formatting (95963ff)
• Improve: Extended uint8 interfaces for Java, Swift, ObjC, JS (82e1196)
• Improve: Smaller & faster striped_locks_gt (4a47d18)
• Docs: SIFT & BigANN subsets (1c60c6c)
• Fix: ObjC compilation issue for b1x8 (d54383d)
• Improve: Reuse eval.random_vectors in scripts (2b7f469)
• Improve: New benchmarking suite (4f6d2d7)
• Improve: Type annotations for Python SDK & scripts (c39b15b)
• Improve: Extended bench.cpp & datasets for 100M scale (dac569c)
• Make: Split flaky CIBW jobs & add retries (5e0139f)
• Make: Drop deprecated SPDX license expression (#744) (145272f)
• Make: Drop EOL language versions and standardize CI runners (0a803a9)
• Fix: Don't compile NumKong C sources into macOS wheel (2c44fb0)
• Make: Bump deps for MSVC compatibility (36981b0)
• Fix: Re-generate JNI header for hardware_acceleration checks (1a3a86e)
• Make: Newer /Zc:preprocessor for MSVC in binding.gyp (79cd39f)
• Fix: NumKong dispatch for Sorensen and binary vector dimensions (5546dd6)
• Fix: Use RAND_MAX instead of INT_MAX for random test vectors (d45a616)
• Make: Use "Trusted Publishing" for NPM (8a02676)
• Fix: PyTest module collision & PeachPy on Python 3.11+ (8ed2548)
• Fix: MSVC type-conversion warnings (#709) (98486cd)
• Improve: Test remove/rename/multi in Rust (17dfac3)
• Make: Arm32-friendly NumKong version (6f65872)
• Fix: Move-safe init for Rust SDK (#704) (05b5eb0)
• Fix: Check for config.expansion early in add (fb2e2f7)
• Fix: Stale vector references in refine_ within index_get::update (#731) (451e2d5)
• Fix: Deduplication logic in form_reverse_links_ (#729) (9ea8233)
• Make: Switch to Ruff linter (#737) (bff98ed)
• Fix: OpenMP flags forwarding in build.rs (#724) (46256a2)
• Fix: Heap-buffer-overflow in HNSW search via sorted_buffer_gt (5697c60)
• Fix: Serial Float16 conversion fallbacks (49b8b2c)
• Make: Avoid npm ci in non-dependency-locked state (42e9203)
• Make: Bump Clang due to LLVM optimizer bug (5bd15bb)
• Make: Don't version removed dependency lock files (62724d4)
• Fix: Hash set iterator returns deleted tombstones in equal_range (#727) (7d95f6f)
• Fix: MAP_FAILED on Windows (#742) (786a7d9)
• Fix: Escalate MAP_FAILED on POSIX (#722) (ba47347)
• Fix: Strip nested sub-directories from package distributions (#733) (3832c02)
• Fix: Guard Float16 APIs with #if arch(arm64) in Swift SDK (#739) (5bcf787)
• Docs: Rust Cos metric docstring (#734) (c9ff5b8)
• Fix: Avoid IntPtr global out-of-scope alias in C# (5413341)
• Improve: Expose & test Float8 in Java, ObjC, Go, JS, Rust (bebb539)
• Fix: Stale symbol name in join demo (ef0dc4f)
• Improve: Allow non-sequential vector IDs in bench (11bcde1)
• Make: Pull NumKong from PyPi (1b46d1b)
• Make: Pull NumKong via CMake (b0f8708)
• Make: Pull SwiftPM NumKong dependency (eede286)
• Make: Pull NumKong crate (46347ec)
• Make: Switch from SimSIMD v6 to NumKong v7 submodule (b2361c3)
• Docs: Link capitalization (98eeb9e)
• Fix: Set equality comparison in Numba JIT PyTest (30381b7)
• Fix: Avoid duplicate neighbor slots in HNSW reverse links (#699) (6995246)
• Fix: Postfix operators for member/candidate iterators in index_gt (#718) (7fbf086)
• Fix: Potential UB in index_dense::add (#721) (6b69e95)
• Fix: Accurately track total allocated memory in memory_mapping_allocator_gt (#608) (3c25f20)
• Fix: Python serialized_length missing argument (#714) (5cde3fc)