Paper · 2026 · Open access

Winnex Madhava Direct: Production-Grade Deterministic Vector Search Benchmark

Klenio Araujo Padilha

Zenodo (CERN European Organization for Nuclear Research) · Published 28 June 2026 · DOI 10.5281/zenodo.21011419

Open full textCite this paper

In brief

“Winnex Madhava Direct: Production-Grade Deterministic Vector Search Benchmark” is a 2026 paper by Klenio Araujo Padilha, published in Zenodo (CERN European Organization for Nuclear Research). OpenAlex counts 0 citations as of 16 November 2026. It is open access: the full text is free to read at Zenodo (CERN European Organization for Nuclear Research).

Citations
0
Authors
1

Abstract

# Winnex Madhava Direct: Benchmark and Validation ## Abstract The Winnex Madhava Direct is a deterministic, auditable vector search method that replaces brute-force cosine similarity with progressive dimensionality reduction via QR-orthogonalized Johnson-Lindenstrauss projections on the Stiefel manifold V_k(R^d). Unlike HNSW or IVF — which rely on non-deterministic graph heuristics — Madhava Direct provides mathematically transparent search with known quality bounds. ## Key Results Benchmarked on News Category Dataset (10K documents, all-MiniLM-L6-v2 384D via QJL to 128D, GPU P100): | Method | NDCG@10 | Time(ms) | vs Exact | NDCG/ms | |--------|---------|----------|----------|---------| | FAISS FlatIP (exact) | 0.4610 | 0.439 | 100% | 1.050 | | FAISS HNSW(ef=64) | 0.4692 | 0.299 | 101.8% | 1.569 | | Madhava Direct (Default) | 0.3504 | 0.732 | 76.0% | 0.479 | | Madhava Direct (Conservative) | 0.4055 | 1.001 | 88.0% | 0.405 | | Madhava Direct (Maximum) | 0.4242 | 1.384 | 92.0% | 0.307 | ## Scaling Performance | Corpus Size | FlatIP(ms) | HNSW(ms) | Madhava(ms) | Speedup vs FlatIP | |------------|-----------|---------|------------|------------------| | 1,000 | 0.863 | 2.256 | 0.395 | 2.2x | | 5,000 | 13.252 | 34.816 | 0.540 | 24.5x | | 10,000 | 30.279 | 43.344 | 1.042 | 29.0x | 50,000 367.932 435.429 3.259 112.9x 100,000 1209.688 1258.467 6.921 174.8x ## Performance Profile by Pool Configuration Configuration Pool Sizes NDCG@10 Time(ms) Ops |-------------|-----------|---------|----------|-----| Minimal 500->100->25 0.2791 (60.5%) 0.839 11,600 Default 1000->200->50 0.3504 (76.0%) 0.857 16,800 Conservative 2000->400->100 0.4055 (88.0%) 1.001 27,200 Maximum 5000->1000->200 0.4242 (92.0%) 1.384 55,200 ## Technical Contributions 1. **QR-Orthogonalized Projections**: Ensures P^T P = I_k (Stiefel manifold), making Pythagorean decomposition valid for Cauchy-Schwarz bounds. 2. **Progressive Dimensionality**: 4D -> 16D -> 64D -> exact 128D cosine, with each stage filtering candidates before the next. 3. **Memory Predictability**: O(N * sum(dims)) = 40MB for 100K documents, vs ~500MB+ for HNSW graph. 4. **Deterministic Execution**: Same query, same data, same result. Unlike HNSW random graph construction. 5. **Build Efficiency**: 19.1ms for 10K documents, vs 200ms+ for HNSW graph construction. ## Complexity Analysis - **Search Complexity**: O(N * d_1 + K_1 * d_2 + K_2 * d_3 + K_3 * D), where d_i are projection dimensions and K_i are pool sizes. The first stage (4D) operates on all N documents, giving O(N * 4). Subsequent stages operate on constant-sized pools. - **Build Complexity**: O(N * D * sum(dims)) for initial projections, done once. - **Memory**: O(N * D + N * sum(dims)) = O(N * (D + sum(dims))). ## Comparison with Prior Work This…

Cite this paper

Formatted from the catalogue record. Check names, capitals and page numbers against the paper before you submit.

APA 7

Padilha, K. A. (2026). Winnex Madhava Direct: Production-Grade Deterministic Vector Search Benchmark. Zenodo (CERN European Organization for Nuclear Research). https://doi.org/10.5281/zenodo.21011419

MLA 9

Padilha, Klenio Araujo. “Winnex Madhava Direct: Production-Grade Deterministic Vector Search Benchmark.” Zenodo (CERN European Organization for Nuclear Research), 2026. https://doi.org/10.5281/zenodo.21011419.

Chicago

Padilha, Klenio Araujo. 2026. “Winnex Madhava Direct: Production-Grade Deterministic Vector Search Benchmark.” Zenodo (CERN European Organization for Nuclear Research). https://doi.org/10.5281/zenodo.21011419.

Harvard

Padilha, K. A. (2026) ‘Winnex Madhava Direct: Production-Grade Deterministic Vector Search Benchmark’, Zenodo (CERN European Organization for Nuclear Research). doi:10.5281/zenodo.21011419.

IEEE

K. A. Padilha, “Winnex Madhava Direct: Production-Grade Deterministic Vector Search Benchmark,” Zenodo (CERN European Organization for Nuclear Research), 2026, doi: 10.5281/zenodo.21011419.

BibTeX

@article{padilha2026winnex,
  title     = {{Winnex Madhava Direct: Production-Grade Deterministic Vector Search Benchmark}},
  author    = {Padilha, Klenio Araujo},
  journal   = {Zenodo (CERN European Organization for Nuclear Research)},
  year      = {2026},
  doi       = {10.5281/zenodo.21011419},
  url       = {https://doi.org/10.5281/zenodo.21011419}
}

Download for a reference manager: BibTeX (.bib) · RIS for Zotero, Mendeley and EndNote (.ris)

Topic and keywords

Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Image and Video Retrieval Techniques

  • vector search
  • deterministic search
  • Stiefel manifold
  • dimensionality reduction
  • cosine similarity
  • HNSW
  • in vitro fertilization
  • NDCG

Work with this paper

Most-cited papers in Computer Science

All highly cited Computer Science papers

Questions about this paper

How do I cite this paper in APA style?

Padilha, K. A. (2026). Winnex Madhava Direct: Production-Grade Deterministic Vector Search Benchmark. Zenodo (CERN European Organization for Nuclear Research). https://doi.org/10.5281/zenodo.21011419

How many times has this paper been cited?

0 times, according to OpenAlex on 16 November 2026. Google Scholar and Scopus count different sets of sources, so their numbers differ.

Is this paper free to read?

Yes. It is open access at Zenodo (CERN European Organization for Nuclear Research): sign in and open the full text from this page.

Who wrote this paper?

Klenio Araujo Padilha.

What is the DOI of this paper?

10.5281/zenodo.21011419, permanently at https://doi.org/10.5281/zenodo.21011419. Sign in and open it from this page.

Bibliographic data from OpenAlex (CC0), last checked 16 November 2026. ResearchGuru does not host this paper; once you are signed in, the links open it at the publisher or the repository that does.