Uh oh!
There was an error while loading. Please reload this page.
chore(deps): bump actions/github-script from 7 to 9 - #3
Merged
ajianaz merged 1 commit intoAug 28, 2026
Merged
Conversation
|
dependabotBotforce-pushed
the
dependabot/github_actions/actions/github-script-9
branch
from
August 28, 2026 04:09
dcb0a81 to
1914830Compareajianaz
commented
Aug 28, 2026
Contributor
@dependabot rebase |
ContributorAuthor
Looks like this PR is already up-to-date with develop! If you'd still like to recreate it from scratch, overwriting any edits, you can request |
🔍 Cora AI Code ReviewReview powered by cora-code · BYOK · MIT |
ajianaz
commented
Aug 28, 2026
Contributor
@dependabot rebase |
Bumps [actions/github-script](https://github.com/actions/github-script) from 7 to 9. - [Release notes](https://github.com/actions/github-script/releases) - [Commits](actions/github-script@v7...v9) --- updated-dependencies: - dependency-name: actions/github-script dependency-version: '9' dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] <support@github.com>
dependabotBotforce-pushed
the
dependabot/github_actions/actions/github-script-9
branch
from
August 28, 2026 04:22
1914830 to
9160b7bCompareajianaz
commented
Aug 28, 2026
Contributor
Merge-gate triage: Cora Review MAJOR finding ( |
Uh oh!
There was an error while loading. Please reload this page.
ajianaz added a commit
that referenced
this pull request
Aug 28, 2026
* chore(cora): raise review llm max_tokens to 16384 glm-5.3 is a reasoning model; on dependency-lockfile diffs the default 4096-token budget is exhausted during reasoning (finish_reason=length, content=null), which made the CI cora-review fallback SARIF fire on PRs #2/#3. 16384 leaves room for reasoning plus the verdict. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> * chore: retrigger CI Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> --------- Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com>
ajianaz added a commit
that referenced
this pull request
Aug 28, 2026
* chore: align .cora.yaml with current cora schema (#4) * chore: align .cora.yaml with current cora schema The old file used `review.enabled`, which the current cora binary rejects (unknown field), silently disabling pre-commit review. Mirrors the uteke .cora.yaml structure with vecq-specific rules (determinism, zero-dependency core). * fix: gate NEON-only bindings behind cfg(aarch64) bpv and q_rot are only used on the NEON path; on x86_64 builds they were unused variables, failing clippy -D warnings in CI. * fix: restore bpv under cfg(aarch64), verify local clippy clean * fix: gate NEON test bindings behind cfg(aarch64) for x86_64 builds * fix: cfg-gate remaining NEON test bindings * fix: cfg-gate the NEON parity test at function level --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * ci: add release workflow (crates.io publish via tag) (#6) * chore: align .cora.yaml with current cora schema (#4) (#5) * chore: align .cora.yaml with current cora schema The old file used `review.enabled`, which the current cora binary rejects (unknown field), silently disabling pre-commit review. Mirrors the uteke .cora.yaml structure with vecq-specific rules (determinism, zero-dependency core). * fix: gate NEON-only bindings behind cfg(aarch64) bpv and q_rot are only used on the NEON path; on x86_64 builds they were unused variables, failing clippy -D warnings in CI. * fix: restore bpv under cfg(aarch64), verify local clippy clean * fix: gate NEON test bindings behind cfg(aarch64) for x86_64 builds * fix: cfg-gate remaining NEON test bindings * fix: cfg-gate the NEON parity test at function level --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * ci: add release workflow (crates.io publish via tag, uteke pattern) --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * chore: bump version to 0.1.1 (#8) * chore: align .cora.yaml with current cora schema (#4) (#5) * chore: align .cora.yaml with current cora schema The old file used `review.enabled`, which the current cora binary rejects (unknown field), silently disabling pre-commit review. Mirrors the uteke .cora.yaml structure with vecq-specific rules (determinism, zero-dependency core). * fix: gate NEON-only bindings behind cfg(aarch64) bpv and q_rot are only used on the NEON path; on x86_64 builds they were unused variables, failing clippy -D warnings in CI. * fix: restore bpv under cfg(aarch64), verify local clippy clean * fix: gate NEON test bindings behind cfg(aarch64) for x86_64 builds * fix: cfg-gate remaining NEON test bindings * fix: cfg-gate the NEON parity test at function level --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * release: v0.1.1 crates.io publish pipeline (#7) * chore: align .cora.yaml with current cora schema (#4) * chore: align .cora.yaml with current cora schema The old file used `review.enabled`, which the current cora binary rejects (unknown field), silently disabling pre-commit review. Mirrors the uteke .cora.yaml structure with vecq-specific rules (determinism, zero-dependency core). * fix: gate NEON-only bindings behind cfg(aarch64) bpv and q_rot are only used on the NEON path; on x86_64 builds they were unused variables, failing clippy -D warnings in CI. * fix: restore bpv under cfg(aarch64), verify local clippy clean * fix: gate NEON test bindings behind cfg(aarch64) for x86_64 builds * fix: cfg-gate remaining NEON test bindings * fix: cfg-gate the NEON parity test at function level --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * ci: add release workflow (crates.io publish via tag) (#6) * chore: align .cora.yaml with current cora schema (#4) (#5) * chore: align .cora.yaml with current cora schema The old file used `review.enabled`, which the current cora binary rejects (unknown field), silently disabling pre-commit review. Mirrors the uteke .cora.yaml structure with vecq-specific rules (determinism, zero-dependency core). * fix: gate NEON-only bindings behind cfg(aarch64) bpv and q_rot are only used on the NEON path; on x86_64 builds they were unused variables, failing clippy -D warnings in CI. * fix: restore bpv under cfg(aarch64), verify local clippy clean * fix: gate NEON test bindings behind cfg(aarch64) for x86_64 builds * fix: cfg-gate remaining NEON test bindings * fix: cfg-gate the NEON parity test at function level --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * ci: add release workflow (crates.io publish via tag, uteke pattern) --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * chore: bump version to 0.1.1 for first crates.io release --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * ci: add cora AI code review workflow (parity with uteke) (#13) * ci: add cora AI code review workflow (parity with uteke) Runs codecoradev/cora-review-action@v1 on every PR targeting develop or main, posting AI review feedback using org CORA_* secrets, matching the uteke repo reference workflow. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> * chore: retrigger CI to verify cora review secrets Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> --------- Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * fix(ci): correct CLA bot exemption to match actual bot logins (#14) dependabot PR authors have user.login `dependabot[bot]`, not `app/dependabot`, so the old skip list never matched and the CLA job failed every dependabot PR. Switch to an exact allowlist of trusted bot logins (dependabot/renovate/github-actions); the legacy app/* forms stay as fallbacks. A substring [bot] match was rejected in review — it would exempt arbitrary third-party GitHub Apps. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * chore(deps): bump rand from 0.8.8 to 0.10.2 (#2) Bumps [rand](https://github.com/rust-random/rand) from 0.8.8 to 0.10.2. - [Release notes](https://github.com/rust-random/rand/releases) - [Changelog](https://github.com/rust-random/rand/blob/master/CHANGELOG.md) - [Commits](rust-random/rand@0.8.8...0.10.2) --- updated-dependencies: - dependency-name: rand dependency-version: 0.10.2 dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * chore(deps): bump actions/github-script from 7 to 9 (#3) Bumps [actions/github-script](https://github.com/actions/github-script) from 7 to 9. - [Release notes](https://github.com/actions/github-script/releases) - [Commits](actions/github-script@v7...v9) --- updated-dependencies: - dependency-name: actions/github-script dependency-version: '9' dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * chore(cora): raise review llm max_tokens to 16384 (#15) * chore(cora): raise review llm max_tokens to 16384 glm-5.3 is a reasoning model; on dependency-lockfile diffs the default 4096-token budget is exhausted during reasoning (finish_reason=length, content=null), which made the CI cora-review fallback SARIF fire on PRs #2/#3. 16384 leaves room for reasoning plus the verdict. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> * chore: retrigger CI Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> --------- Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * feat(core): keyed API — add_keyed/remove_keyed/tombstones (#10) (#16) Adds the keyed layer requested for incremental integrators (issue #10, unblocks the cora-code adoption audit): - add_keyed(key, v): insert-or-replace under a stable u64 key (usearch-style), returns the slot index - remove_keyed(key) -> bool: tombstones the slot; storage stays in place so slot indices remain stable until compact() - search_keyed(q, k): top-k by key, tombstones skipped - compact(): rebuilds in place, dropping tombstones (keys preserved, slot indices shift) - key_of/contains_key/slots/tombstones introspection helpers - len()/is_empty() now report live vectors - to_bytes() writes live slots only: tombstones are dropped on disk without disturbing in-memory slot indices. File format unchanged (v1.1); keys are an in-memory concern — persist key->slot mappings in your own metadata store. NEON4 batch scoring is untouched: tombstoned slots are still scored to keep batches dense, then filtered before entering the top-k heap, preserving bit-identical results. TDD: 7 new tests cover keyed search, replace, stability across remove/serialize, compaction, drained indexes, keyed adds after a file load, and dim-mismatch panics. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * feat(core): x86_64 AVX2 scoring path (parity with NEON, bit-identity tested) (#17) * feat(core): x86_64 AVX2 scoring path, bit-identical to scalar (#11) Adds an explicit AVX2 nibble-gather scoring path mirroring the NEON one, selected at runtime with is_x86_feature_detected! (AVX2 is not baseline on x86_64): - score_avx2: 8 lanes = 8 scalar buckets; per 8-byte block, lane j computes q[2b]*lut[lo_b] + q[2b+1]*lut[hi_b] (mul, mul, add — no FMA) and accumulates in block order, then the same pairwise-tree reduction — structurally identical to the scalar loop, so results are bit-identical - score_avx24: 4-vector batch sharing the q deinterleave - LUT gather via vgatherdps (scale 4, nibble indices); q even/odd deinterleave via shuffle_ps + vpermps fixup - search() dispatches to the AVX2 batch when available; score() dispatches per call Bit-identity is enforced by new tests (avx2_matches_scalar_bitwise, avx24_matches_avx2_bitwise, search dispatch == scalar) including a dim-8 case covering the 4-byte tail. The x86_64 CI job runs these on native AVX2 hardware (Rosetta hides AVX2 via CPUID, so local emulated runs cannot exercise the path). docs/BENCHMARK.md: per-architecture scoring-path table; x86_64 numbers explicitly pending native measurement, Rosetta excluded. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> * fix(clippy): iterate acc via enumerate in score_avx24 batch loop clippy 1.98 (CI stable) flags the 0..4 indexing loop under needless_range_loop; local 1.89 does not. Same semantics. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> --------- Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * docs: canonical SQLite BLOB storage pattern (#12) (#18) Adds docs/SQLITE.md: when to embed the index vs standalone file, the single-row and per-shard schema shapes, save/load with rusqlite in one transaction, atomicity notes (WAL + synchronous=NORMAL — no manual checkpoint choreography needed for correctness), measured save/load latencies at 1k/10k/50k vectors (dim 768, ~514 B/vec), and pitfalls (memory duplication on load, slot stability across save/reload interacting with the keyed API, churn-heavy workloads, embed size threshold, seed discipline). README links the page and mentions the per-architecture scoring-path matrix in BENCHMARK.md. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * chore: bump version to 0.2.0 (#19) Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * fix: resolve Cargo.lock conflict markers left from merge The merge committed conflict markers into Cargo.lock; restore the develop-side lock (vecq-core 0.2.0) and verify the workspace builds. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> --------- Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>
5 tasks
ajianaz added a commit
that referenced
this pull request
Aug 30, 2026
* chore: align .cora.yaml with current cora schema (#4) * chore: align .cora.yaml with current cora schema The old file used `review.enabled`, which the current cora binary rejects (unknown field), silently disabling pre-commit review. Mirrors the uteke .cora.yaml structure with vecq-specific rules (determinism, zero-dependency core). * fix: gate NEON-only bindings behind cfg(aarch64) bpv and q_rot are only used on the NEON path; on x86_64 builds they were unused variables, failing clippy -D warnings in CI. * fix: restore bpv under cfg(aarch64), verify local clippy clean * fix: gate NEON test bindings behind cfg(aarch64) for x86_64 builds * fix: cfg-gate remaining NEON test bindings * fix: cfg-gate the NEON parity test at function level --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * ci: add release workflow (crates.io publish via tag) (#6) * chore: align .cora.yaml with current cora schema (#4) (#5) * chore: align .cora.yaml with current cora schema The old file used `review.enabled`, which the current cora binary rejects (unknown field), silently disabling pre-commit review. Mirrors the uteke .cora.yaml structure with vecq-specific rules (determinism, zero-dependency core). * fix: gate NEON-only bindings behind cfg(aarch64) bpv and q_rot are only used on the NEON path; on x86_64 builds they were unused variables, failing clippy -D warnings in CI. * fix: restore bpv under cfg(aarch64), verify local clippy clean * fix: gate NEON test bindings behind cfg(aarch64) for x86_64 builds * fix: cfg-gate remaining NEON test bindings * fix: cfg-gate the NEON parity test at function level --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * ci: add release workflow (crates.io publish via tag, uteke pattern) --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * chore: bump version to 0.1.1 (#8) * chore: align .cora.yaml with current cora schema (#4) (#5) * chore: align .cora.yaml with current cora schema The old file used `review.enabled`, which the current cora binary rejects (unknown field), silently disabling pre-commit review. Mirrors the uteke .cora.yaml structure with vecq-specific rules (determinism, zero-dependency core). * fix: gate NEON-only bindings behind cfg(aarch64) bpv and q_rot are only used on the NEON path; on x86_64 builds they were unused variables, failing clippy -D warnings in CI. * fix: restore bpv under cfg(aarch64), verify local clippy clean * fix: gate NEON test bindings behind cfg(aarch64) for x86_64 builds * fix: cfg-gate remaining NEON test bindings * fix: cfg-gate the NEON parity test at function level --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * release: v0.1.1 crates.io publish pipeline (#7) * chore: align .cora.yaml with current cora schema (#4) * chore: align .cora.yaml with current cora schema The old file used `review.enabled`, which the current cora binary rejects (unknown field), silently disabling pre-commit review. Mirrors the uteke .cora.yaml structure with vecq-specific rules (determinism, zero-dependency core). * fix: gate NEON-only bindings behind cfg(aarch64) bpv and q_rot are only used on the NEON path; on x86_64 builds they were unused variables, failing clippy -D warnings in CI. * fix: restore bpv under cfg(aarch64), verify local clippy clean * fix: gate NEON test bindings behind cfg(aarch64) for x86_64 builds * fix: cfg-gate remaining NEON test bindings * fix: cfg-gate the NEON parity test at function level --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * ci: add release workflow (crates.io publish via tag) (#6) * chore: align .cora.yaml with current cora schema (#4) (#5) * chore: align .cora.yaml with current cora schema The old file used `review.enabled`, which the current cora binary rejects (unknown field), silently disabling pre-commit review. Mirrors the uteke .cora.yaml structure with vecq-specific rules (determinism, zero-dependency core). * fix: gate NEON-only bindings behind cfg(aarch64) bpv and q_rot are only used on the NEON path; on x86_64 builds they were unused variables, failing clippy -D warnings in CI. * fix: restore bpv under cfg(aarch64), verify local clippy clean * fix: gate NEON test bindings behind cfg(aarch64) for x86_64 builds * fix: cfg-gate remaining NEON test bindings * fix: cfg-gate the NEON parity test at function level --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * ci: add release workflow (crates.io publish via tag, uteke pattern) --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * chore: bump version to 0.1.1 for first crates.io release --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * ci: add cora AI code review workflow (parity with uteke) (#13) * ci: add cora AI code review workflow (parity with uteke) Runs codecoradev/cora-review-action@v1 on every PR targeting develop or main, posting AI review feedback using org CORA_* secrets, matching the uteke repo reference workflow. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> * chore: retrigger CI to verify cora review secrets Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> --------- Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * fix(ci): correct CLA bot exemption to match actual bot logins (#14) dependabot PR authors have user.login `dependabot[bot]`, not `app/dependabot`, so the old skip list never matched and the CLA job failed every dependabot PR. Switch to an exact allowlist of trusted bot logins (dependabot/renovate/github-actions); the legacy app/* forms stay as fallbacks. A substring [bot] match was rejected in review — it would exempt arbitrary third-party GitHub Apps. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * chore(deps): bump rand from 0.8.8 to 0.10.2 (#2) Bumps [rand](https://github.com/rust-random/rand) from 0.8.8 to 0.10.2. - [Release notes](https://github.com/rust-random/rand/releases) - [Changelog](https://github.com/rust-random/rand/blob/master/CHANGELOG.md) - [Commits](rust-random/rand@0.8.8...0.10.2) --- updated-dependencies: - dependency-name: rand dependency-version: 0.10.2 dependency-type: direct:production update-type: version-update:semver-minor ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * chore(deps): bump actions/github-script from 7 to 9 (#3) Bumps [actions/github-script](https://github.com/actions/github-script) from 7 to 9. - [Release notes](https://github.com/actions/github-script/releases) - [Commits](actions/github-script@v7...v9) --- updated-dependencies: - dependency-name: actions/github-script dependency-version: '9' dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * chore(cora): raise review llm max_tokens to 16384 (#15) * chore(cora): raise review llm max_tokens to 16384 glm-5.3 is a reasoning model; on dependency-lockfile diffs the default 4096-token budget is exhausted during reasoning (finish_reason=length, content=null), which made the CI cora-review fallback SARIF fire on PRs #2/#3. 16384 leaves room for reasoning plus the verdict. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> * chore: retrigger CI Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> --------- Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * feat(core): keyed API — add_keyed/remove_keyed/tombstones (#10) (#16) Adds the keyed layer requested for incremental integrators (issue #10, unblocks the cora-code adoption audit): - add_keyed(key, v): insert-or-replace under a stable u64 key (usearch-style), returns the slot index - remove_keyed(key) -> bool: tombstones the slot; storage stays in place so slot indices remain stable until compact() - search_keyed(q, k): top-k by key, tombstones skipped - compact(): rebuilds in place, dropping tombstones (keys preserved, slot indices shift) - key_of/contains_key/slots/tombstones introspection helpers - len()/is_empty() now report live vectors - to_bytes() writes live slots only: tombstones are dropped on disk without disturbing in-memory slot indices. File format unchanged (v1.1); keys are an in-memory concern — persist key->slot mappings in your own metadata store. NEON4 batch scoring is untouched: tombstoned slots are still scored to keep batches dense, then filtered before entering the top-k heap, preserving bit-identical results. TDD: 7 new tests cover keyed search, replace, stability across remove/serialize, compaction, drained indexes, keyed adds after a file load, and dim-mismatch panics. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * feat(core): x86_64 AVX2 scoring path (parity with NEON, bit-identity tested) (#17) * feat(core): x86_64 AVX2 scoring path, bit-identical to scalar (#11) Adds an explicit AVX2 nibble-gather scoring path mirroring the NEON one, selected at runtime with is_x86_feature_detected! (AVX2 is not baseline on x86_64): - score_avx2: 8 lanes = 8 scalar buckets; per 8-byte block, lane j computes q[2b]*lut[lo_b] + q[2b+1]*lut[hi_b] (mul, mul, add — no FMA) and accumulates in block order, then the same pairwise-tree reduction — structurally identical to the scalar loop, so results are bit-identical - score_avx24: 4-vector batch sharing the q deinterleave - LUT gather via vgatherdps (scale 4, nibble indices); q even/odd deinterleave via shuffle_ps + vpermps fixup - search() dispatches to the AVX2 batch when available; score() dispatches per call Bit-identity is enforced by new tests (avx2_matches_scalar_bitwise, avx24_matches_avx2_bitwise, search dispatch == scalar) including a dim-8 case covering the 4-byte tail. The x86_64 CI job runs these on native AVX2 hardware (Rosetta hides AVX2 via CPUID, so local emulated runs cannot exercise the path). docs/BENCHMARK.md: per-architecture scoring-path table; x86_64 numbers explicitly pending native measurement, Rosetta excluded. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> * fix(clippy): iterate acc via enumerate in score_avx24 batch loop clippy 1.98 (CI stable) flags the 0..4 indexing loop under needless_range_loop; local 1.89 does not. Same semantics. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> --------- Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * docs: canonical SQLite BLOB storage pattern (#12) (#18) Adds docs/SQLITE.md: when to embed the index vs standalone file, the single-row and per-shard schema shapes, save/load with rusqlite in one transaction, atomicity notes (WAL + synchronous=NORMAL — no manual checkpoint choreography needed for correctness), measured save/load latencies at 1k/10k/50k vectors (dim 768, ~514 B/vec), and pitfalls (memory duplication on load, slot stability across save/reload interacting with the keyed API, churn-heavy workloads, embed size threshold, seed discipline). README links the page and mentions the per-architecture scoring-path matrix in BENCHMARK.md. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * chore: bump version to 0.2.0 (#19) Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * feat(core): Matryoshka-aware working_dim truncation (#30) * feat(core): Matryoshka-aware working_dim truncation (#24) Adds VecqIndex::with_working_dim(dim, working_dim, seed): quantize only the leading working_dim dimensions of dim-dimensional vectors, for Matryoshka-trained embedding models (EmbeddingGemma, text-embedding-3, ...). Truncation happens BEFORE normalization and rotation — RHDH mixes dimensions, so post-rotation truncation would not be Matryoshka-equivalent. - Callers always pass full-dim vectors/queries; the index truncates internally, keeping the add/search call sites unchanged - working_dim() accessor; dim() remains the caller-facing dim - Format v1.2: the previously-reserved u16 header field carries working_dim (0 = full dim); payload layout identical to v1.1. Readers accept v1, v1.1 and v1.2; writers emit v1.2 - README: Matryoshka recipe; BENCHMARK.md format note Tests (TDD, 6 new): working_dim == dim is bit-identical to new(); synthetic-Matryoshka data keeps neighbor ranking at ~1/3 storage; v1.2 round-trip preserves working_dim; legacy v1.1 files load bit-identically; working_dim > dim and 0 panic. Validated locally across repeated runs: native workspace suites 2x, x86_64 target, clippy -D warnings both targets, fmt, bench regression check. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> * fix(core): reject working_dim > u16::MAX unless equal to dim Cora CI review caught a real wrap: to_bytes casts working_dim to u16, so working_dim=70000 with dim=100000 wrote 4464 — a file that parses successfully with the wrong code layout (silent wrong search results), and working_dim=65536 wrapped to 0 (read as full dim). Construction now rejects working_dim > u16::MAX unless it equals dim (stored as 0 in the header). Regression tests for the panic case and for full-dim indexes with dim > u16::MAX. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> --------- Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * feat(core): keyed parity with usearch — relabel + multi-vectors-per-key (#26) (#33) - relabel(old_key, new_key) -> bool: renames a key in place (slots untouched); rejects unknown source / taken target, self-relabel is a successful no-op - add_keyed_multi(key, v): appends another vector under one key (usearch's multi mode); keys hold 1..n slots - remove_keyed(key): removes every slot of the key; remove_keyed_at(key, slot): removes one slot, shrinking the key back to the single-slot representation when one slot remains - search_keyed dedupes multi-slot keys (one entry per key, best slot's score), so results can hold fewer than k entries - compact rebuilds both single- and multi-key mappings Internals: single-slot keys stay in a HashMap<u64, usize>; multi keys move to HashMap<u64, Vec<usize>> — the common case remains allocation-free. append_slot no longer registers keys (caller-owned), fixing a double-map registration bug the tests caught in the promote path. Tests (TDD, 4 new): relabel semantics, multi add/search/dedupe, remove/remove_at lifecycle, multi replace + relabel + compact round trip. README usage updated. Validated locally per repo discipline: native workspace 5/5 x5 runs, x86_64 target 41/41, clippy -D warnings 0 findings both targets, fmt clean, bench numbers unchanged. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * bench: head-to-head vs TurboQuant and RaBitQ at 4 bits (#28) (#34) Adds bin/vs_quantizers: identical synthetic clustered datasets, exact f32 cosine ground truth, measuring recall@10 / ms-per-query / build time / bytes-per-vector for vecq vs TurboQuant-MSE 4-bit vs RaBitQ 4-bit brute force (rabitq-rs). TurboQuant is scanned SDC-style through a 16x16 product table built from OUR codebook instance (TurboQuantMSE::with_codebook) so the LUT is structurally guaranteed to match the encoder; the crate exposes no ADC path or rotation accessor. RaBitQ runs via its first-class BruteForceRabitqIndex, as implemented (internal rayon during train). Results (aarch64, single-threaded, recall deterministic across runs): vecq leads scan speed (2.1-2.7 vs 15.0-15.7 / 37.9-39.3 ms/q at n=10k dim=768) and recall at dim 768 (0.840 vs 0.798 / 0.827), with the honest caveat that competitors use ~25% less storage per vector (no per-vector scale, no power-of-two padding). Full tables + methodology notes + go/no-go for #22/#23 in docs/BENCHMARK.md. Competitor crates are bench-only dependencies; vecq-core stays zero-dependency. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * docs: add 0.2.0 changelog entry (#31) (#35) Backfills the v0.2.0 section (keyed API, AVX2 path, Matryoshka working_dim, SQLite guide, competitor benchmark, format v1.2, CLA bot fix) so the tag's user-facing changes are visible. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * fix(core): persist the keyed map — file format v1.3 (#32) (#36) Keyed state used to vanish on save/reload: from_bytes rebuilt codes and scales but never the key->slot map, so search_keyed/remove_keyed were dead on every reloaded index (user-filed repro in #32). Format v1.3 appends a keyed-slot table after the codes block: [entries u32][slot u32 + key u64 each]. Slot ids are dense positions among the serialized (alive) slots — matching the reader's slot space — not in-memory indices, which can exceed the serialized count when tombstones are dropped (caught by the multi-key round-trip test). Entries are validated (slot bounds, duplicate keys) and restore_keys rebuilds both the single- and multi-key maps, so reloaded indexes are fully keyed-capable (search_keyed, remove_keyed(_at), relabel, compact all work). Readers accept v1/v1.1/v1.2/v1.3; writers emit v1.3. docs (SQLITE.md, README, BENCHMARK.md) updated; CHANGELOG Unreleased entry added. Tests (TDD, 2 new): the issue's exact repro (keys survive reload, reloaded index remains keyed-capable) and a multi-key round trip with remove_keyed_at + compact + re-serialization stability. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * feat(core): cascade search — 2-bit prefilter + 4-bit rescore (#22) (#37) Adds an opt-in approximate search mode: enable_cascade() derives a 2-bit signature per slot from the stored nibbles (the two high bits, i.e. a coarse Lloyd re-quantization — padded/4 bytes per slot), and search_cascade(q, k, r) ranks slots by L1 distance between the query's and each slot's 2-bit codes (pure integer math, 16-entry pair LUT), keeps the r closest, and rescores those with the standard bit-identical 4-bit path. Design notes vs the original issue sketch: - 2-bit L1 instead of 1-bit Hamming: the Lloyd-Max codebook is not sign-symmetric (MSB of the code is a +0.164 threshold, not zero), so plain sign-bit Hamming degenerates on clustered data (measured 6/10 top-10 overlap at r=50 on n=1k). The 2-bit re-quantization keeps the derivation trivially consistent between database (nibble >> 2) and query (quantize_4bit >> 2 on the unnormalized rotated dims, via the new PreparedQuery::rnorm) while being strictly more discriminative (same measurement: 6/10 at r=50 but 1.00 at r=200; monotone). - No format change: signatures are derived from codes, not persisted — enable_cascade() must be called after from_bytes; add/replace/compact drop the signatures (documented). - Determinism: integer prefilter + tie-break toward smaller slots + bit-identical rescore; r >= live is exactly search() (tested bitwise). Tombstoned slots are skipped. Tests (TDD, 7 new): bitwise equivalence at r >= n, monotone recall in r on both a moderate and an adversarial clustered set, tombstone skipping + determinism, reload + compact behavior, empty index, r clamping, working_dim composition. README cascade recipe added. Local validation (repeated): native workspace 6/6 x5 runs, x86_64 50/50, clippy -D warnings 0/0, fmt clean, bench unchanged (0.35-0.38 ms/q baseline). Pre-commit cora review: no findings. Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * feat(core): residual quantization — opt-in recall mode, exact-norm two-term scoring, format v1.4 (#23) (#38) - Encode: second Lloyd-Max pass over the first-pass residual (format v1.4, readers accept v1-v3)\n- Scoring: two-term ADC with the EXACT reconstruction norm (cross term 2*rms*<d0,d1> + true <d1,d1> energy folded in at encode time; no format change, kernels untouched)\n- Real EmbeddingGemma: recall@1 0.990 / recall@10 0.984 at ~3x compression, 1.43x scan cost (plain: 0.910/0.958 @ 5.98x; usearch HNSW 0.995 @ 1x)\n- Adversarial clustered set: residual 0.87 vs plain 0.66 (was 0.58 with the approximate denominator)\n- real_residual_validation.rs (ignored, dataset-gated) + docs/BENCHMARK.md section\n- Re-tested & re-benchmarked before merge per policy: fmt/clippy clean, 54/54 tests debug+release, plain-path recall bit-identical baseline across real/vs_usearch/vs_quantizers * feat(core): configurable Lloyd-Max width (4/5/6-bit), default 5-bit sweet spot (#39) (#41) * feat(core): configurable Lloyd-Max width (4/5/6-bit), default 5-bit sweet spot (#39) - set_bits() width option, frozen after first add; default 5-bit (4.79x, recall@10 0.979 on real EmbeddingGemma) per locked design - LSB-first bit-packing with 16-bit RMW window (codes span byte boundaries at 5/6-bit); pack_code/unpack_code shared by encode and reference scoring - score_wide_scalar: portable per-width scoring via lloyd::centroids; 4-bit NEON path untouched (bit-identical, byte-identical files) - format v1.5: explicit width byte after header (plain non-4-bit only); 4-bit plain stays v1.3, residual v1.4; readers accept all - cascade search asserts 4-bit (signatures are nibble-based) - widths bench bin: end-to-end 4/5/6-bit + residual verification - tests: width matrix (B/vec, v1.3/v1.4/v1.5, round-trips), naive score parity at every width, wide-vs-residual recall dominance - README: configurable width, updated guarantees and format list * fix(test): AVX2 dispatch parity test covers all widths (#39) The x86_64-only test still referenced the 4-bit nibble path while the default width is now 5-bit; loop 4/5/6 with the width-matched scalar reference. Compile-checked for x86_64-unknown-linux-gnu. * perf(neon): batch-4 wide kernel for 5/6-bit scoring (#40) - score_neon_wide4: one u64 window per 8 dims replaces 8 unpack_code calls; centroids gathered through L1; NEON mul/add with per-lane bucket order identical to score_wide_scalar (bit-identity tested bitwise for 5/6-bit x dims {128,384}, incl. batch boundaries) - search dispatch: 4-bit LUT path and wide path share the batch loop; residual rescore unchanged (4-bit only) - x86_64 unchanged (scalar wide path; AVX2 wide kernel = follow-up) - E2E (real, d=768, n=2000): 5-bit 4.27 -> 3.23 ms/q; extraction is now the bottleneck — split-layout (nibble stream + vqtbl blend) is the documented next lever in #40 --------- Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * docs: benchmark width matrix (#39/#40) + changelog for Unreleased (#42) - BENCHMARK.md: results table gains the 5-bit default row; new width matrix section (4/5/6-bit + residual, post wide-kernel); analysis and limitations updated for the extraction-bound 5/6-bit path with the #40 ceiling analysis; conclusion reflects configurable width - CHANGELOG: Unreleased gains #23/#39/#40 entries Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * feat(core): zero-copy VecqView — serve an index straight from bytes (#25) (#43) - VecqView::from_bytes borrows v1.2+ file bytes (mmap, Box<[u8]>, any owner); no payload copying; keyed map absent by design, views are dense (to_bytes drops tombstones) - shared kernel dispatch extracted (score_raw_dispatch, score_batch4) so index and view hit identical kernels in identical order — bitwise parity tested at 4/5/6-bit, residual, and working_dim, plus malformed-input rejection incl. trailing keyed-table extent - view search uses the same batch-4 loop shape as the index (parity: 19.96 vs 19.85 ms/q warm at n=12k) - mmap cold-start harness (bench-only memmap2 dep; core stays zero-dependency): map+parse 64 us vs 4.9 ms full load (~76x faster time-to-ready, 12k vectors, 5-bit, 7.7 MB file) - docs/SQLITE.md: when to skip the BLOB — mmap view guidance Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * docs: README overhaul for the width/view era + stale-claim sweep (#44) - README: Modes table (4/5/6-bit + residual), zero-copy VecqView usage, Persistence & serving section, honest Performance paragraph for the default width, corrected Status (crates.io prepared, not published), Why-vecq table moved to default-width numbers - lib.rs module docs: configurable width, LSB-first packing, VecqView - BENCHMARK.md: dataset bullet names format v1.5 + 5-bit default - SQLITE.md: width-aware size formula, refreshed size table, VecqView pointer in the size section - CHANGELOG: #25 entry Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> * chore: bump version to 0.3.0 (#45) Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> --------- Signed-off-by: ajianaz <ajianaz@users.noreply.github.com> Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: ajianaz <ajianaz@users.noreply.github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.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 freeto join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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.
Bumps actions/github-script from 7 to 9.
Release notes
Sourced from actions/github-script's releases.
... (truncated)
Commits
3a2844bMerge pull request #700 from actions/salmanmkc/expose-getoctokit + prepare re...ca10bbdfix: use@octokit/core/types import for v7 compatibility86e48e2merge: incorporate main branch changesc108472chore: rebuild dist for v9 upgrade and getOctokit factoryafff112Merge pull request #712 from actions/salmanmkc/deployment-false + fix user-ag...ff8117eci: fix user-agent test to handle orchestration ID81c6b78ci: use deployment: false to suppress deployment noise from integration tests3953cafdocs: update README examples from@v8to@v9, add getOctokit docs and v9 brea...c17d55bci: add getOctokit integration test joba047196test: add getOctokit integration tests via callAsyncFunction