Add PiPNN: high-performance alternative index builder (7.6-10x faster graph build)

[SeliMeli]

Summary

Adds PiPNN (Pick-in-Partitions Nearest Neighbors, arXiv:2602.21247) as an alternative graph construction algorithm for DiskANN's disk index.

PiPNN replaces Vamana's sequential greedy-insert with batch GEMM-based construction:

1. Partition — Random Ball Carving splits dataset into overlapping clusters
2. Leaf Build — All-pairs distance via BLAS GEMM within each cluster, extract bidirectional k-NN
3. HashPrune — LSH-based reservoir merges edges from overlapping partitions
4. Final Prune — Optional RobustPrune pass (same alpha as Vamana)

This makes the build embarrassingly parallel with no sequential point-to-point dependency.

Benchmark Results (8 threads, 1M points)

Build Performance

	Mimir QG 1M (384d, cosine)	Mimir QG 1M SQ1 (1-bit)	SIFT 1M (128d, L2)
Graph build speedup	10x (201s → 20s)	7.6x (136s → 18s)	9.7x (73s → 7.5s)
End-to-end speedup	5.4x (222s → 41s)	4.2x (157s → 38s)	6.3x (88s → 14s)

Search Quality (same disk index format — search path is identical)

Dataset	Metric	Vamana Recall	PiPNN Recall	Gap	Vamana QPS	PiPNN QPS
Mimir QG 1M	R@1000	96.19%	94.74%	-1.45%	15.6	15.0
Mimir QG 1M SQ1	R@1000	95.75%	90.15%	-5.60%	15.3	15.9
SIFT 1M (L=100)	R@10	99.50%	98.53%	-0.97%	258	159
SIFT 1M (L=200)	R@10	99.85%	99.59%	-0.26%	153	147

Recall gap is tunable via parameters (replicas, l_max, leaf_k). SQ1 gap needs further tuning.

Memory

Peak build RSS is ~1.8 GB vs Vamana's ~1.2 GB (Mimir QG 1M FP). Optimized from initial 3.2 GB via:

• f16-native build (generic over T: VectorRepr, converts on-the-fly)
• Smaller partition GEMM stripes (16 MB target)
• Consuming extract_graph(self) to free reservoirs during extraction
• malloc_trim between build phases

Changes

• New crate: diskann-pipnn (~5700 lines) — builder, partition, leaf_build, hash_prune, gemm, quantize
• diskann-disk — PiPNN integration into disk index build pipeline, build phase timing output
• diskann-benchmark — build_algorithm: PiPNN config support
• Feature-gated behind pipnn — zero impact on existing Vamana builds

Test plan

• 102 unit tests in diskann-pipnn pass
• Mimir QG 1M FP benchmark: recall 94.74%, no regression vs prior runs
• Mimir QG 1M SQ1 benchmark: recall 90.15%
• SIFT 1M benchmark: recall 98.53% (L=100), 99.59% (L=200)
• Build time verified across 3 datasets, 3 configurations
• Peak RSS profiled with /usr/bin/time -v and /proc/self/status VmHWM

🤖 Generated with Claude Code

[microsoft-github-policy-service]

@SeliMeli please read the following Contributor License Agreement(CLA). If you agree with the CLA, please reply with the following information.

@microsoft-github-policy-service agree [company="{your company}"]

Options:

• (default - no company specified) I have sole ownership of intellectual property rights to my Submissions and I am not making Submissions in the course of work for my employer.

@microsoft-github-policy-service agree

• (when company given) I am making Submissions in the course of work for my employer (or my employer has intellectual property rights in my Submissions by contract or applicable law). I have permission from my employer to make Submissions and enter into this Agreement on behalf of my employer. By signing below, the defined term “You” includes me and my employer.

@microsoft-github-policy-service agree company="Microsoft"

Contributor License Agreement

Contribution License Agreement

This Contribution License Agreement (“Agreement”) is agreed to by the party signing below (“You”),
and conveys certain license rights to Microsoft Corporation and its affiliates (“Microsoft”) for Your
contributions to Microsoft open source projects. This Agreement is effective as of the latest signature
date below.

1. Definitions.
   “Code” means the computer software code, whether in human-readable or machine-executable form,
   that is delivered by You to Microsoft under this Agreement.
   “Project” means any of the projects owned or managed by Microsoft and offered under a license
   approved by the Open Source Initiative (www.opensource.org).
   “Submit” is the act of uploading, submitting, transmitting, or distributing code or other content to any
   Project, including but not limited to communication on electronic mailing lists, source code control
   systems, and issue tracking systems that are managed by, or on behalf of, the Project for the purpose of
   discussing and improving that Project, but excluding communication that is conspicuously marked or
   otherwise designated in writing by You as “Not a Submission.”
   “Submission” means the Code and any other copyrightable material Submitted by You, including any
   associated comments and documentation.
2. Your Submission. You must agree to the terms of this Agreement before making a Submission to any
   Project. This Agreement covers any and all Submissions that You, now or in the future (except as
   described in Section 4 below), Submit to any Project.
3. Originality of Work. You represent that each of Your Submissions is entirely Your original work.
   Should You wish to Submit materials that are not Your original work, You may Submit them separately
   to the Project if You (a) retain all copyright and license information that was in the materials as You
   received them, (b) in the description accompanying Your Submission, include the phrase “Submission
   containing materials of a third party:” followed by the names of the third party and any licenses or other
   restrictions of which You are aware, and (c) follow any other instructions in the Project’s written
   guidelines concerning Submissions.
4. Your Employer. References to “employer” in this Agreement include Your employer or anyone else
   for whom You are acting in making Your Submission, e.g. as a contractor, vendor, or agent. If Your
   Submission is made in the course of Your work for an employer or Your employer has intellectual
   property rights in Your Submission by contract or applicable law, You must secure permission from Your
   employer to make the Submission before signing this Agreement. In that case, the term “You” in this
   Agreement will refer to You and the employer collectively. If You change employers in the future and
   desire to Submit additional Submissions for the new employer, then You agree to sign a new Agreement
   and secure permission from the new employer before Submitting those Submissions.
5. Licenses.

• Copyright License. You grant Microsoft, and those who receive the Submission directly or
  indirectly from Microsoft, a perpetual, worldwide, non-exclusive, royalty-free, irrevocable license in the
  Submission to reproduce, prepare derivative works of, publicly display, publicly perform, and distribute
  the Submission and such derivative works, and to sublicense any or all of the foregoing rights to third
  parties.
• Patent License. You grant Microsoft, and those who receive the Submission directly or
  indirectly from Microsoft, a perpetual, worldwide, non-exclusive, royalty-free, irrevocable license under
  Your patent claims that are necessarily infringed by the Submission or the combination of the
  Submission with the Project to which it was Submitted to make, have made, use, offer to sell, sell and
  import or otherwise dispose of the Submission alone or with the Project.
• Other Rights Reserved. Each party reserves all rights not expressly granted in this Agreement.
  No additional licenses or rights whatsoever (including, without limitation, any implied licenses) are
  granted by implication, exhaustion, estoppel or otherwise.

6. Representations and Warranties. You represent that You are legally entitled to grant the above
   licenses. You represent that each of Your Submissions is entirely Your original work (except as You may
   have disclosed under Section 3). You represent that You have secured permission from Your employer to
   make the Submission in cases where Your Submission is made in the course of Your work for Your
   employer or Your employer has intellectual property rights in Your Submission by contract or applicable
   law. If You are signing this Agreement on behalf of Your employer, You represent and warrant that You
   have the necessary authority to bind the listed employer to the obligations contained in this Agreement.
   You are not expected to provide support for Your Submission, unless You choose to do so. UNLESS
   REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING, AND EXCEPT FOR THE WARRANTIES
   EXPRESSLY STATED IN SECTIONS 3, 4, AND 6, THE SUBMISSION PROVIDED UNDER THIS AGREEMENT IS
   PROVIDED WITHOUT WARRANTY OF ANY KIND, INCLUDING, BUT NOT LIMITED TO, ANY WARRANTY OF
   NONINFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
7. Notice to Microsoft. You agree to notify Microsoft in writing of any facts or circumstances of which
   You later become aware that would make Your representations in this Agreement inaccurate in any
   respect.
8. Information about Submissions. You agree that contributions to Projects and information about
   contributions may be maintained indefinitely and disclosed publicly, including Your name and other
   information that You submit with Your Submission.
9. Governing Law/Jurisdiction. This Agreement is governed by the laws of the State of Washington, and
   the parties consent to exclusive jurisdiction and venue in the federal courts sitting in King County,
   Washington, unless no federal subject matter jurisdiction exists, in which case the parties consent to
   exclusive jurisdiction and venue in the Superior Court of King County, Washington. The parties waive all
   defenses of lack of personal jurisdiction and forum non-conveniens.
10. Entire Agreement/Assignment. This Agreement is the entire agreement between the parties, and
    supersedes any and all prior agreements, understandings or communications, written or oral, between
    the parties relating to the subject matter hereof. This Agreement may be assigned by Microsoft.

[Copilot]

Pull request overview

Adds PiPNN as a new (feature-gated) disk-index graph construction algorithm, integrating it into the disk index build pipeline and benchmark/tooling so builders can choose between Vamana and PiPNN.

Changes:

• Introduce new diskann-pipnn crate implementing partitioning, leaf-build, HashPrune merging, and optional final pruning.
• Add BuildAlgorithm to disk build parameters/config and wire it through diskann-disk, diskann-tools, and diskann-benchmark.
• Add build-phase timing output and PiPNN-specific 1-bit SQ reuse path in the disk index builder.

Reviewed changes

Copilot reviewed 24 out of 25 changed files in this pull request and generated 12 comments.

Show a summary per file

File	Description
diskann-tools/src/utils/build_disk_index.rs	Adds build_algorithm to tool-level disk index build parameters and passes it into disk build configuration.
diskann-tools/Cargo.toml	Enables diskann-disk’s pipnn feature for tools.
diskann-providers/src/model/graph/provider/async_/inmem/scalar.rs	Exposes quantizer() accessor to reuse trained SQ params for PiPNN.
diskann-pipnn/src/lib.rs	Defines PiPNN public API, config, error types, and metric serde.
diskann-pipnn/src/builder.rs	Implements end-to-end PiPNN build orchestration + optional final prune + graph save format.
diskann-pipnn/src/partition.rs	Implements RBC partitioning (float + quantized variants) and associated tests.
diskann-pipnn/src/leaf_build.rs	Implements GEMM-based all-pairs distances, kNN extraction, and edge materialization.
diskann-pipnn/src/hash_prune.rs	Implements LSH sketching + per-point reservoir HashPrune and extraction.
diskann-pipnn/src/gemm.rs	Wraps diskann-linalg GEMM calls used by PiPNN.
diskann-pipnn/src/quantize.rs	Implements 1-bit quantization + Hamming-distance helpers and tests.
diskann-pipnn/benches/pipnn_bench.rs	Adds Criterion benchmarks for PiPNN hot paths.
diskann-pipnn/README.md	Documents PiPNN usage, parameters, and expected performance.
diskann-pipnn/Cargo.toml	Adds the new crate manifest + deps.
diskann-disk/src/lib.rs	Re-exports BuildAlgorithm from build configuration.
diskann-disk/src/build/mod.rs	Re-exports BuildAlgorithm for consumers.
diskann-disk/src/build/configuration/mod.rs	Adds build_algorithm module and re-export.
diskann-disk/src/build/configuration/disk_index_build_parameter.rs	Stores selected build algorithm in disk build parameters and exposes accessor.
diskann-disk/src/build/configuration/build_algorithm.rs	Defines BuildAlgorithm enum and PiPNN parameter mapping (feature-gated).
diskann-disk/src/build/builder/build.rs	Integrates PiPNN into disk index build pipeline + phase timing output.
diskann-disk/Cargo.toml	Adds optional dependency + feature flag for diskann-pipnn and test deps.
diskann-benchmark/src/inputs/disk.rs	Adds build_algorithm to disk-index build job input model.
diskann-benchmark/src/backend/disk_index/build.rs	Passes build_algorithm into disk build parameters.
diskann-benchmark/Cargo.toml	Enables diskann-disk/pipnn and adds diskann-pipnn dependency.
Cargo.toml	Adds diskann-pipnn to workspace and workspace deps.
Cargo.lock	Records new crate and dependency graph changes.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

brute_force_knn uses select_nth_unstable_by(actual_k, ...) and then truncates to actual_k. As with the other quickselect usage, pivoting at actual_k can exclude one of the true k-smallest elements (notably for k=1). Pivot at actual_k - 1 when actual_k > 0 before truncating, then sort/truncate.

Suggested change

	if actual_k < dists.len() {
	dists.select_nth_unstable_by(actual_k, |a, b| {
	a.1.partial_cmp(&b.1).unwrap_or(std::cmp::Ordering::Equal)
	if actual_k > 0 && actual_k < dists.len() {
	dists.select_nth_unstable_by(actual_k - 1, |a, b| {
	a.1
	.partial_cmp(&b.1)
	.unwrap_or(std::cmp::Ordering::Equal)

[Copilot]

buf.select_nth_unstable_by(num_assign, ...) partitions around the (num_assign)-th element, but the code then reads buf[0..num_assign]. This can miss one of the true num_assign nearest leaders (e.g., fanout=1 should pivot at 0). Pivot at num_assign - 1 (when num_assign > 0) before taking the first num_assign entries.

[Copilot]

The benchmark input tuple is (m, n, k), but the closure destructures it as &(m, k, _). This drops n and re-binds k to the second element, so the GEMM dimensions passed to sgemm_abt are incorrect and the code likely won't compile due to unused/mismatched bindings. Destructure as &(m, n, k) and pass the correct values.

[Copilot]

gemm::set_blas_threads(1) is called here, but diskann_pipnn::gemm (src/gemm.rs) does not define set_blas_threads. This will fail to compile under --all-targets (benches included). Either remove these calls or add an implementation (likely in diskann-linalg) behind an appropriate feature/target gate.

Suggested change

gemm::set_blas_threads(1);

[Copilot]

leaf_build::build_leaf takes a Metric as its last argument (see src/leaf_build.rs), but the benchmark passes false. This will not compile and also makes the benchmark ambiguous about the distance metric being measured. Pass an explicit diskann_vector::distance::Metric value instead.

[Copilot]

diskann-pipnn is added as an unconditional dependency, but it is not referenced from this crate (and diskann-disk is already optional behind the disk-index feature). This forces PiPNN to compile even when disk-index is disabled, increasing build times/binary size. Consider removing it or making it conditional (e.g., only pulled in via diskann-disk's pipnn feature under disk-index).

Suggested change

diskann-pipnn = { workspace = true }

[Copilot]

select_nth_unstable_by(actual_k, ...) uses actual_k as the pivot index, but the code then reads indices[0..actual_k]. For k-nearest selection this can omit the true minimums (e.g., k=1 should pivot at 0, not 1). Use actual_k - 1 as the pivot (when actual_k > 0) and then take the first actual_k indices (optionally sorting that prefix if order matters).

[Copilot]

Same quickselect issue as above: buf.select_nth_unstable_by(num_assign, ...) followed by reading buf[0..num_assign] does not guarantee selecting the num_assign closest leaders. Use pivot index num_assign - 1 (when > 0), then take the first num_assign entries.

[Copilot]

build_algorithm() returns &BuildAlgorithm, but this if let BuildAlgorithm::PiPNN { .. } = ... pattern matches a by-value BuildAlgorithm and will not compile. Match on a reference (e.g., if matches!(self.disk_build_param.build_algorithm(), &BuildAlgorithm::PiPNN { .. })) or deref to a value if you change the accessor to return an owned value.

Suggested change

	if let BuildAlgorithm::PiPNN { .. } = self.disk_build_param.build_algorithm() {
	if matches!(
	self.disk_build_param.build_algorithm(),
	&BuildAlgorithm::PiPNN { .. }
	) {

[Copilot]

Similarly, this matches!(self.disk_build_param.build_algorithm(), BuildAlgorithm::Vamana) is matching &BuildAlgorithm against a by-value pattern and will not compile under #[cfg(not(feature = "pipnn"))]. The pattern needs to match a reference (e.g., &BuildAlgorithm::Vamana).

Suggested change

	BuildAlgorithm::Vamana
	&BuildAlgorithm::Vamana

[harsha-simhadri]

This is great! and thanks for the pointer.

On the index side: Could you share the benchmark jsons for diskann baseline so we can replicate. As Mark suggests, the baseline seems slower than expected.

On the query side: Please run queries on SSD. While we have a lot to learn from this paper, I want to be careful about disk based graphs. Diskann graphs are designed to have lower diameter which are better for SSD. We would want to see the diameter of graphs PiPNN generated. You could for example measure the number of points that are "n" hops from start for n=1,2,3..10, by fixing degree. A SSD friendly index would have more nodes closer to origin.

The prune method is very interesting, we should consider overall what we can learn and put in core diskann prune and how to make it streaming.