0.1.7 — Optimized CUDA MaxSim + RROQ158 kernel + rename to colsearch + bugfixes

Highlights

This release ships three things at once:

1. Optimized CUDA kernels for MaxSim and RROQ158 on H100

• Fused single-pass mma.sync.b1.b1.s32.and.popc kernel for RROQ158
• Multi-tier (32/64/128/256/512) padded MaxSim dispatcher
• Autotune-aware warmup at production batch sizes
• int64-pointer fix in the Triton MaxSim kernel
• GPU full-corpus fast-path that bypasses LEMUR routing when the preloaded corpus already pays the VRAM cost
• Persistent query/corpus scratch buffers

Result on BEIR-8 (H100): 3.12× geomean QPS over FastPlaid at fp16 (446.7 vs 143.2 QPS), 2.06× at rroq158_gs128 (294.4 QPS). See benchmarks/competitive_benchmark.md.

2. Rename voyager-index → colsearch

• PyPI: pip install colsearch
• GitHub: ddickmann/colsearch (the old ddickmann/voyager-index URL auto-redirects)
• Python package: colsearch (with a voyager_index compat shim)
• Console script: colsearch-server (voyager-index-server retained as an alias)

3. Bugfixes

• Bench disk-tightening: os.rename / os.link instead of shutil.copytree (~100 GB peak disk saved on BEIR-8 sweep).
• CPU whole-corpus fast-path that bypasses the per-query 522k-row numpy fancy-index gather (was making quora rroq158/cpu allocate ~5 GB/query and hang for 90+ minutes).
• RROQ158 hot-path alloc-churn fix that eliminated a 1.2 GB/call churn forcing CUDA allocator GC every ~200 queries.
• Reference-api Dockerfile installs Rust via rustup (pinned 1.94.1) instead of Debian's apt rustc 1.85, which predated the avx512_target_feature stabilization (1.89+) used by the AVX-512 VPOPCNTDQ tier — unblocks docker-smoke.
• Release pipeline now uses skip-existing: true on PyPI publish so colsearch can ship without forcing a no-op version bump on the independently-versioned native crates.

Compatibility (rename)

• pip install voyager-index users: install colsearch instead. Legacy distribution stops at 0.1.6.
• import voyager_index keeps working in 0.1.7 via a thin shim that emits a single DeprecationWarning and eagerly aliases every voyager_index.X.Y submodule to the canonical colsearch.X.Y module in sys.modules so isinstance and enum identity continue to hold. Removed in 0.2.0.
• voyager-index-server console script aliased to colsearch-server for the 0.1.x cycle, removed in 0.2.0.
• VOYAGER_INDEX_PATH env var still honoured but emits a deprecation warning. Migrate to COLSEARCH_INDEX_PATH.
• VOYAGER_BENCH_CPU_TIME_BUDGET_S is honoured; COLSEARCH_BENCH_CPU_TIME_BUDGET_S is the new canonical name.
• Other VOYAGER_* env vars (VOYAGER_RROQ158_N_THREADS, VOYAGER_RROQ158_USE_B1_FUSED) are unchanged in 0.1.7 to avoid breaking ops scripts.
• Docker image: latence/colsearch (was latence/voyager-zero); reference Dockerfile builds colsearch:latest.
• Kubernetes manifests under deploy/k8s/ ship the new colsearch namespace and labels — adjust your overlays.

Native packages

The native crates (latence-shard-engine, latence-solver) are versioned independently of the root colsearch package and only need a republish when their Rust source changes. 0.1.7 is a metadata-only change for them; the published wheels stay at 0.1.6.

Provenance

Historical artefacts under reports/, validation-reports/, research/, notebooks/, and benchmarks/_smoke_*.py / benchmarks/_diag_*.py retain their original voyager-index 0.1.x labels and SHAs. Run identifiers in reports/fast_plaid_head_to_head/results_v7.jsonl (e.g. voyager_fp16, voyager_rroq158_gs128) and the corresponding benchmarks/fast_plaid_head_to_head.py --libraries flags are kept stable so historical JSONL stays readable; they map 1:1 to the colsearch fp16 / rroq158_gs128 lanes.

Install

pip install colsearch

Migrating from voyager-index:

pip uninstall voyager-index
pip install colsearch
# `import voyager_index` keeps working until 0.2.0; update to `import colsearch` at your leisure.

See the full CHANGELOG.md entry for 0.1.7.

0.1.6 — RROQ158 SOTA Default at group_size=128

This release promotes the dim-aware Rroq158Config(group_size=128) lane to
the build-time default for newly created RROQ158 indexes — closing the
production-validation arc started in 0.1.5 (Phase 7 / Phase 8).

Highlights

• ~13% smaller per-token storage (~40 vs ~46 bytes/token at dim=128;
  ~6.4× smaller than fp16, up from ~5.5×).
• CPU p95 ~10–30% faster on the BEIR-6 mean (one fewer scale load per
  group in the popcount kernel; nfcorpus −22%, scifact −15%, scidocs −10%,
  fiqa within +2% noise).
• NDCG@10 within ±0.005 of the previous gs=32 baseline; per-dataset
  mean Δ vs gs=32 across BEIR-6 = +0.0006 — Pareto-equal in quality
  while delivering smaller storage AND lower-or-equal CPU p95 on every
  dataset measured.
• Dim-aware fallback _resolve_group_size(requested, dim)
  transparently steps down to gs=64 / gs=32 (with a log warning) on
  production corpora whose token dim is not divisible by 128, so the new
  default works on dim=64 / 96 / 128 / 160 alike without caller changes.

Headline BEIR-6 averages (rroq158 gs=128 default)

Codec	NDCG@10	R@100	GPU p95 (ms)	CPU p95 (ms)
fp16 (baseline)	0.5206	0.7360	4.0 (1.00×)	103 (1.00×)
rroq158 (gs=128, default)	0.5069	0.7298	4.8 (1.20×)	310 (3.00×)
rroq158 (gs=32)	0.5063	0.7282	4.8 (1.20×)	325 (3.15×)
rroq4_riem	0.5158	0.7345	8.5 (2.13×)	580 (5.63×)

Full per-dataset / per-codec sweep:
reports/beir_2026q2_gs128/.

Migration

Existing on-disk indexes are unaffected — the manifest carries the
build-time group_size and only newly built indexes pick up the new
default.

• Pin Rroq158Config(group_size=32) to restore the previous default
  exactly.
• Pin Rroq158Config(group_size=64) for the safest cross-dataset choice
  (covers high-intra-token-variance corpora like arguana).

See docs/guides/quantization-tuning.md
for the full decision matrix and per-dim recipe table.

Install

pip install --upgrade voyager-index==0.1.6
# native (optional, for the Rust SIMD CPU kernel + tabu solver):
pip install --upgrade voyager-index[native]==0.1.6
# full (server + multimodal + native + GPU triton):
pip install --upgrade voyager-index[full,gpu]==0.1.6

Full changelog and merged PRs in
CHANGELOG.md.

0.1.5

Changelog

This changelog tracks the official shipped OSS release line. Older draft notes
that did not correspond to a published release were removed so version history
reads in release order again.

Unreleased

0.1.5 — Release Gate Hotfix

This release republishes the shard-engine decomposition work on a clean CI line
after fixing the small lint regressions that slipped through the initial 0.1.4
cut.

Release integrity

• fixed the shard refactor parity script bootstrap so the release lint lane
  accepts the repo-local import setup
• normalized import ordering and explicit public exports in the refactor-touching
  files that failed the hosted Ruff gate
• bumped the root package and supported native packages onto the 0.1.5 line
  so the hotfix release cleanly supersedes the drafted 0.1.4 cut

0.1.4 — Shard Engine Decomposition And Release Evidence

This release keeps the shard product surface stable while decomposing the large
shard-engine modules behind compatibility facades and hardening the parity
evidence required to ship that refactor safely.

Shard engine maintainability

• split the shard manager, store, fetch pipeline, LEMUR router, builder, WAL,
  and ColBANDIT reranker into focused internal modules while preserving public
  import paths
• reduced config coupling by separating serving configuration from sweep-only
  configuration behind compatibility exports
• introduced internal protocols for router, store, fetch, reranker, and native
  exact backends to narrow cross-module ownership

Runtime capability visibility

• surfaced fallback and capability state for LEMUR routing, pinned staging, and
  native exact execution through shard statistics and reference API metadata
• added startup logging for shard capability selection so development and
  production runs expose fallback decisions explicitly

Validation and release confidence

• added shard refactor contract coverage for import compatibility, artifact
  parity, query trace stability, and runtime capability reporting
• added a machine-readable shard refactor parity report and wired it into CI so
  release evidence is reproducible instead of ad hoc
• bumped the root package and supported native packages onto the 0.1.4 line
• refreshed release hygiene checks to validate the aligned package versions

0.1.3 — Production Release Hardening

This release closes the gap between the public product story and the shipped
package, native-wheel, and release pipeline surfaces.

Packaging and install surface

• added a canonical voyager-index[full] install profile for the full public CPU-safe surface
• added shard-native and broadened native so the public native story now covers both latence-shard-engine and latence-solver
• bumped the root package and supported native packages onto the 0.1.3 line
• tightened package data so the shipped sdist includes the graph quality fixture required by release validation

Graph-aware production path

• kept latence-graph as a public optional extra and pinned it to the verified public latence>=0.1.1 line
• clarified throughout the docs that the graph lane can consume compatible prebuilt graph data directly and remains additive to the shard-first hot path
• preserved the graph route-conformance, provenance, and retrieval-uplift evidence as a distinct proof layer from shard performance benchmarks

CI, release, and OSS hygiene

• expanded the native release bundle to include the shard-engine wheel alongside the solver wheel
• tightened release documentation and automation around clean-install rehearsal, native-wheel validation, and publish gating
• refreshed the README, install docs, issue templates, and contributor guidance around the supported production lane
• added repo-governance files for dependency updates, code ownership, and contributor conduct

0.1.2 — Shard Production Surface

This release makes the shard engine the clear public product surface.

Retrieval and serving

• production-wired shard search with LEMUR routing, ColBANDIT, and Triton MaxSim
• shard scoring controls exposed for int8, fp8, and roq4
• durable CRUD, WAL, checkpoint, recovery, and shard admin endpoints
• multi-worker single-host reference server posture

API and SDK

• base64 vector transport helpers exposed from voyager_index.transport
• public HTTP API accepts base64 payloads for dense and multivector requests
• shard configuration knobs surfaced on collection create, search, and info APIs
• dense hybrid mode selection documented and shipped as rrf or tabu

Docs and DX

• README, quickstart, API docs, and top-level guides rewritten around the shard-first story
• benchmark methodology documented with a 100k comparison placeholder table
• reference API examples now lead with base64 and shard-friendly install profiles

Release and packaging

• release notes and changelog chronology cleaned up
• CI trimmed to shard-only production lanes plus solver validation
• supported native add-on story reduced to latence_solver

0.1.0 — Initial OSS Foundation Release

Initial public package release for voyager-index.

Foundation

• installable voyager_index package and published OSS packaging surface
• durable reference FastAPI service
• dense, late-interaction, and multimodal collection kinds
• CRUD, restart-safe persistence, and public examples

Retrieval

• exact MaxSim exports through the public package
• CPU-safe MaxSim fallback when Triton is unavailable
• hybrid dense + BM25 retrieval
• optional solver-backed refinement via latence_solver

Multimodal

• preprocessing helpers for renderable source documents
• multimodal model registry and provider seams
• ColPali-oriented multimodal retrieval surface

v0.1.3

Production release hardening for the full public voyager-index surface.

• adds the canonical voyager-index[full] install profile for the full CPU-safe production lane
• publishes both supported native packages on the 0.1.3 line: latence-shard-engine and latence-solver
• keeps the Latence graph lane optional, additive, and policy-driven on top of the shard-first hot path
• gates publishing behind the full release validation stack, including clean-install, graph-route, solver, shard, and Docker checks

v0.1.2

Prebuilt Rust wheels for Linux and macOS.

Install

pip install voyager-index                # pure Python
pip install voyager-index[native]        # + prebuilt Rust kernels (Linux x86_64, macOS x86_64/arm64)
pip install voyager-index[native,server] # + FastAPI reference server

Platforms

• Linux x86_64 (manylinux 2.28)
• macOS Intel (x86_64)
• macOS Apple Silicon (arm64)
• Python 3.10, 3.11, 3.12