Repository navigation
native: Highway is the default vector provider where the wheel ships its library - #154
Merged
Merged
Conversation
…its library (#153) The engine forces jvector's scalar provider in every native image unless ARCADEDB_VECTOR_PROVIDER says otherwise. The loader now sets it to native (Highway) when libjvector-<arch>.so sits next to the engine library, which the wheel ships on Linux x86-64 and aarch64; a value the user set wins, and default forces the scalar provider. Measured in #153 (timed laptop A/B): index builds 2.2 to 2.5 times faster and queries 14 to 30 percent faster at 768 and 1536 dimensions, the same build and 6 percent slower queries at 128, identical recall. Docs: which-package gets a Vector search row (JVM faster to build and query, native smaller in memory), vectors.adoc a section on the kernels. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… section (#153) Numbers generated from the vector-wheels A/B score (docs_numbers.py), numcheck clean. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…n the JVM backend, like test_native_heap.py) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #153
Default. The engine forces jvector's scalar provider in every native image unless
ARCADEDB_VECTOR_PROVIDERsays otherwise. The loader (_native/lib.py) now sets it tonative(jvector's Highway kernels) whenlibjvector-<arch>.sosits next to the engine library, which the wheel ships on Linux x86_64 and aarch64. A value the user set wins;ARCADEDB_VECTOR_PROVIDER=defaultforces the scalar provider. macOS and Windows keep the scalar provider (no library shipped). No native-image change: the variable was already read byArcadeNative.Why (timed laptop A/B in #153, medians of per-round ratios, 100,000 vectors,
26.11.1.dev0): against the scalar provider, Highway builds indexes 2.2 to 2.5 times faster and answers queries 14 to 30 percent faster at 768 and 1536 dimensions; at 128 dimensions the build takes the same time and queries are 6 percent slower. Recall@10 is identical for all three arms. Who it serves: users of the defaultarcadedb(native) package with vector indexes, most embeddings being 384 to 1536 dimensions.Docs.
which-package.adocgets a Vector search row: with Highway, native index builds take 1.28 to 1.62 times as long as on the JVM build and median queries 1.22 to 1.87 times as long, while native peak memory is 0.51 to 0.67 of the JVM build's.how-to/vectors.adocgets a section on the distance kernels of each package and the variable. All numbers are generated from the A/B score.Tests (
tests/test_native_vector_provider.py): the helper picks Highway when the library is present, a user value wins (default,panama,native), no library keeps scalar; in a fresh process, the provider the isolate runs follows the shipped library, anddefaultforcesDefaultVectorizationProvider.🤖 Generated with Claude Code