Third-Party and Source Acknowledgements

The original kodama-cpp code is released under the MIT License. Identified third-party portions retain their compatible upstream terms. The formal component and snapshot audit is in PROVENANCE.md.

Maintainer-authored source lineage

Stefano Cacciatore confirmed that tkcaccia <tkcaccia@gmail.com> is his historical Git identity, that he developed KODAMA and fastPLS, and that his adapted contributions in kodama-cpp are released under MIT. His publication contact is stefano.cacciatore@icgeb.org.

The public KODAMA package is GPL (>= 2), and the public fastPLS package is GPL-3. Local ports of the KODAMA optimization procedure and fastPLS SIMPLS/LDA implementation are separately relicensed under MIT by Stefano Cacciatore for this repository. This permission covers only material whose copyright he owns; package and publication coauthors remain credited.

src/passing_message.cpp reproduces the mathematics of the maintainer-authored passing.message function in tkcaccia/KODAMAextra and replaces its R-level nearest-neighbor call and accumulation loops with the package-owned exact grid search and float32 C++ kernels. KODAMAextra remains credited as the source lineage; this preprocessing operation is not claimed as a new method.

src/spatial_features.cpp, src/spatial_features_cuda.cu, and src/spatial_features_metal.mm are original MIT-licensed implementations of a low-rank spatial covariance screening method. No SPARK or SPARK-X source code is included or linked. SPARK-X is used only as an external benchmark and is credited in the method documentation and benchmark report.

The native visualization, CPU HNSW, and Metal implementation were adapted from the MIT-licensed fastEmbedR snapshot 814350a5ca69b0c26e6df40377636f109055f84b. The exact 2D/3D grid search was adapted from the MIT-licensed faissR snapshot b317a9715dd33ad3a49cf7989a83b4f7f9f7b389. Both repositories were developed and committed by Stefano Cacciatore under the Git identities documented above.

FAISS HNSW

The package-owned CPU HNSW implementation in src/native_knn.cpp was distilled from the HNSW organization in FAISS 1.14.3, commit 0ca9df4792b173d573044ee14ca0704780176e82 (MIT). The file retains the Meta Platforms and Stefano Cacciatore notices. The complete FAISS license is in licenses/FAISS-LICENSE.

Metal KNN

src/metal_backend.mm is adapted from fastEmbedR and fastPLS. Its KNN organization was informed by FAISS 1.14.3 and contains portions inherited from the fused IVF list-scan/top-k organization of MLXPorts/Faiss-mlx commit d092af559375144fc719cd88a10e414f92c625fa. Those portions remain Apache-2.0; the FAISS-derived and kodama-cpp portions remain MIT. The file is therefore marked MIT AND Apache-2.0 and carries Meta Platforms, Sydney Bach / The Solace Project, and Stefano Cacciatore notices. The complete licenses are in licenses/FAISS-LICENSE and licenses/FAISS-MLX-LICENSE.

The inspected Faiss-mlx Git history uses the author identity Sydney Renee, whereas the fastEmbedR source notice records Sydney Bach, The Solace Project. The inherited fastEmbedR notice is retained pending clarification upstream.

Native CUDA search

The package-owned CUDA exact search, IVF-Flat search, and k-means code in src/native_cuda_backend.cu is informed by the algorithmic organization of FAISS 1.14.3 (MIT), RAPIDS cuVS commit ad9e2d2a617c8d51e3eebc920e5a60ad8dc59bcd (Apache-2.0), and the native Metal work in fastEmbedR. Its k-means initialization, seed convention, Lloyd iteration semantics, and empty-cluster repair are adapted from FAISS 1.14.3 faiss/Clustering.cpp and random utilities. The source file therefore retains the Meta Platforms notice and MIT terms. No FAISS, cuVS, RAFT, or RMM binary is vendored or linked, and no cuVS source is included. The upstream license texts are retained in licenses/FAISS-LICENSE and licenses/CUVS-LICENSE.

UMAP and openTSNE

src/visualization.cpp, src/embedding_cuda_kernels.cu, and the Metal UMAP and openTSNE optimizers in src/metal_backend.mm are adapted from fastEmbedR's MIT implementation. Its documented mathematical and architectural references include the BSD-licensed UMAP reference implementation, openTSNE, t-SNE-CUDA, Rtsne behavior, opt-SNE/Multicore-opt-SNE, umappp, and ensmallen; AppleSiliconFFT (MIT); and mlx-vis (Apache-2.0). No source from the GPL-licensed uwot package is included in the core. See PROVENANCE.md for the complete classification and the rule governing future adaptations. The current port includes fastEmbedR's direct binary/fuzzy CSR construction, smooth-kNN bandwidth calculation, backend-specific CPU CSR/CUDA COO-CSR/Metal clean-row UMAP schedules, CPU float32 openTSNE optimizer, CUDA embedding kernels, fixed-point atomic Metal UMAP updates, and native Metal FFT-grid openTSNE. The Metal port remains part of the mixed-license src/metal_backend.mm file because that file also contains separately identified Apache-2.0 Faiss-mlx-derived search organization.

PCA

src/pca.cpp, src/pca_cuda.cu, and their public wrappers implement a standalone float32 randomized PCA using the current fastEmbedR/fastPLS strategy for Gaussian sketches, power iterations, backend-specific automatic settings, score/loadings orientation, and explained-variance reporting. The implementation does not link either package, Armadillo, Eigen, or RAFT. The fastPLS-derived strategy is included under Stefano Cacciatore's MIT relicensing declaration above; the complete fastPLS publication and package credit remain required.

Runtime and build dependencies

CPU, Metal, and CUDA numerical builds do not link FAISS, cuVS, fastEmbedR, fastPLS, or an R/Python runtime. Metal builds use Apple system frameworks (Foundation, Metal, and Metal Performance Shaders). CUDA builds use CUDA Toolkit libraries. OpenMP is optional.

No external FAISS or cuVS binary is distributed with this repository. Binary package distributors must include notices required by the particular CUDA, OpenMP, or platform binaries they redistribute.