feat(rust): add search_with_filter to CAGRA Index#2019
Open
jamie8johnson wants to merge 1 commit intorapidsai:mainfrom
Open
feat(rust): add search_with_filter to CAGRA Index#2019jamie8johnson wants to merge 1 commit intorapidsai:mainfrom
jamie8johnson wants to merge 1 commit intorapidsai:mainfrom
Conversation
Author
Author
|
Labels needed: improvement + non-breaking (external contributor, cannot self-label). |
jamie8johnson
added a commit
to jamie8johnson/cqs
that referenced
this pull request
Apr 13, 2026
Override VectorIndex::search_with_filter for CagraIndex: builds a bitset from the predicate on host, uploads to GPU, and passes it to CAGRA's traversal-time filter via search_with_filter (patched cuvs). Eliminates the 3x over-fetch workaround — k=100 goes directly to GPU as k=100 with a bitset, not k=300 unfiltered + post-filter. Better recall for type-filtered and language-filtered queries. Also: - Cargo.toml [patch.crates-io] pointing to ../cuvs-patched (local patched cuvs 26.4 with search_with_filter, upstream PR rapidsai/cuvs#2019) - Demoted CAGRA itopk_size clamp warning to debug level Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
jamie8johnson
added a commit
to jamie8johnson/cqs
that referenced
this pull request
Apr 14, 2026
* feat: enrichment ablation + optimal routing + batch base index support Two-arm eval at 78% summary coverage with per-category SPLADE: - Base (no summaries): 42.3% R@1 - Enriched (with summaries): 41.9% R@1 - Oracle (best per category): 43.8% R@1 (+1.9pp) Router updated based on per-category results: - type_filtered → DenseBase (+8.4pp: 41.7% vs 33.3%) - multi_step → DenseBase (+2.9pp: 23.5% vs 20.6%) - structural/conceptual/cross_language stay enriched Batch handler now supports base/enriched HNSW routing: - Added base_hnsw field + base_vector_index() to BatchContext - dispatch_search classifies queries and routes to base when appropriate - CQS_FORCE_BASE_INDEX=1 env var for eval A/B testing - Fixes daemon always using enriched regardless of classification Other: - Demote CAGRA itopk_size clamp warning to debug level - Fix stale Cargo.toml comment about cuVS CUDA compatibility - Cargo.lock updated for merged dep bumps (#935-#938) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * feat: GPU-native CAGRA filtered search via bitset Override VectorIndex::search_with_filter for CagraIndex: builds a bitset from the predicate on host, uploads to GPU, and passes it to CAGRA's traversal-time filter via search_with_filter (patched cuvs). Eliminates the 3x over-fetch workaround — k=100 goes directly to GPU as k=100 with a bitset, not k=300 unfiltered + post-filter. Better recall for type-filtered and language-filtered queries. Also: - Cargo.toml [patch.crates-io] pointing to ../cuvs-patched (local patched cuvs 26.4 with search_with_filter, upstream PR rapidsai/cuvs#2019) - Demoted CAGRA itopk_size clamp warning to debug level Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: clippy div_ceil + CI-safe cuvs patch via git dep - Fix clippy::manual_div_ceil in CAGRA bitset construction - Switch [patch.crates-io] from local path to git repo (CI can't access ../cuvs-patched) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * refactor: simplify CAGRA to non-consuming search (cuVS 26.4) cuVS 26.4 changed Index::search(self) to search(&self). This eliminates the entire take-rebuild cycle that was causing SIGABRT in the daemon under sustained use (repeated GPU index rebuilds corrupted CUDA state). Removed: - IndexRebuilder RAII guard - rebuild_index_with_resources / ensure_index_rebuilt - dataset field (was cached for rebuilds) - Mutex<Option<Index>> → single Mutex<GpuState> (resources + index) The index is now built once and reused for all searches. search_impl() is shared between filtered and unfiltered paths. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * docs: eval results, tears, roadmap updates - 8 eval run results (enrichment ablation, CAGRA filtering, routing) - Updated roadmap: marked completed items, added HyDE/CAGRA items - Updated tears with session summary and next priorities - Added cuvs-fork-push to .gitignore Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * release: v1.24.0 — GPU-native CAGRA filtering + daemon stability - Bump version 1.23.0 → 1.24.0 - CHANGELOG: add v1.24.0 entry, fold stale [Unreleased] into v1.23.0 - README: note cuVS 26.04 conda requirement + patched crate Highlights: - CAGRA native bitset filtering (GPU-side, replaces 3x over-fetch) - Batch/daemon base index routing fix - Router: type_filtered + multi_step → DenseBase - cuVS 26.4: fixes daemon SIGABRT under sustained CAGRA load - cagra.rs simplified (−357 lines via non-consuming search) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: jamie8johnson <jamie8johnson@users.noreply.github.com> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
3 tasks
Add `Index::search_with_filter()` that accepts a bitset filter via DLPack ManagedTensor. The C API `cuvsCagraSearch()` already supports `cuvsFilter` with BITSET type, but the Rust bindings hardcoded `NO_FILTER`. This exposes the existing capability. The bitset is a 1-D uint32 device tensor with ceil(n_rows / 32) elements. Bit = 1 includes the row, bit = 0 excludes it. Filtering happens during graph traversal, not post-retrieval. Includes test: builds a 256-point index, filters to even-indexed rows, verifies all returned neighbors pass the filter.
jamie8johnson
force-pushed
the
rust-search-filter
branch
from
01c3950 to
006aa84
Compare
jamie8johnson
added a commit
to jamie8johnson/cqs
that referenced
this pull request
Apr 21, 2026
The hook was matching any path containing `/src/`, which included vendored Rust subtrees like `cuvs-fork-push/rust/cuvs/src/`. When an Edit fired against a cuvs file, the hook tried to invoke `python3 .claude/hooks/pre-edit-impact.py` with cwd=cuvs-fork-push, where that path doesn't exist — silently blocking the Edit with a 'No such file' error. Hit during today's rebase of rapidsai/cuvs#2019 to resolve a merge conflict; had to work around by writing the file via a Python script in Bash. Tighten the path predicate: derive cqs root from the script's own location, then accept only files under `<cqs_root>/src/` or `<cqs_root>/tests/`. cuvs-fork-push, evals/, samples/, and any future vendored Rust subtree all silently bail. Verified with four smoke cases: - cuvs path → exit 0, no `cqs impact` invocation - cqs/src path → exit 0, runs `cqs impact` (no output for hypothetical fns) - cqs/tests path → exit 0, runs `cqs impact` - cqs/evals/*.rs → exit 0, no invocation (not under src/ or tests/) Co-authored-by: jamie8johnson <jamie8johnson@users.noreply.github.com> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
aamijar
added
improvement
Member
|
/ok to test 841f0f3 |
@aamijar, there was an error processing your request: See the following link for more information: https://docs.gha-runners.nvidia.com/cpr/e/2/ |
Member
|
/ok to test 006aa84 |
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
2 participants
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.
Summary
Add
Index::search_with_filter()to the CAGRA Rust bindings that accepts a bitset filter via DLPackManagedTensor. The C APIcuvsCagraSearch()already supportscuvsFilterwithBITSETtype, but the Rust bindings hardcodedNO_FILTER. This exposes the existing capability to Rust consumers.Details
The bitset is a 1-D
uint32device tensor withceil(n_rows / 32)elements. Bit = 1 includes the row, bit = 0 excludes it. Filtering happens during CAGRA graph traversal, not post-retrieval, giving better recall than over-fetch-and-filter approaches.The existing
search()method is unchanged (backward compatible).search_with_filter()is additive.Motivation
We use CAGRA via the Rust crate for a code search tool. Currently we post-filter search results by metadata (chunk type, language), which requires 3x over-fetching to compensate for filtered-out candidates. Native bitset filtering would eliminate the over-fetch, reduce GPU work, and improve recall for filtered queries.
Test
test_cagra_search_with_filter: builds a 256-point index, creates a bitset that includes only even-indexed rows, searches with the filter, and verifies all returned neighbors are even-indexed.Changes
rust/cuvs/src/cagra/index.rs: Addedsearch_with_filter()method + test