Your knowledge base is a tree. Stop forcing it through cosine.

Both are models of the same hyperbolic geometry. Lorentz (hyperboloid model) is numerically more stable for high-dimensional embeddings — no points exploding near the boundary of the unit ball. Poincaré is prettier for 2D demos. Lorentz is what you want for 64–128 dim production embeddings. If you specifically need Poincaré coordinates, I can add a conversion (free, exp/log map is one line).

Those are research libraries — great, I read them. HyperspaceDB is my own implementation focused on production use: native Lorentz storage, exp/log maps cached at a base point, batched arccosh distance, Parquet I/O. It's not "better" universally; it's specialized for embedding + retrieving private corpora end-to-end without GPU clusters. If you want to swap in geoopt's optimizer, the math is the same, the parquet output works downstream.

Works, but tell me. Code repos are often deeply hierarchical (modules → packages → classes → methods) — actually a great hyperbolic use case. I'll use a code-aware tokenizer instead of a prose one. Mention the language(s) in your email.

Yes. Tested on Polish, English, mixed PL/EN, and short snippets in German/Russian. Embedding model is multilingual. If your corpus is in a language I haven't tested, I'll say so up front and either run it (probably fine) or refund.

Multilingual sentence-transformer (768-dim) → Riemannian projection onto Lorentz hyperboloid → optimized to dim 64 or 128. Specific base model named in the report. If you have a strong preference (e.g. bge-m3, nomic-embed-text), tell me — I can swap it.

Your corpus stays on my machine for the 48–72h job. Not uploaded to any third-party API for embedding (multilingual model runs locally on Apple Silicon / CUDA). After delivery I delete the source files. I don't reuse your corpus for training. NDA on request — say so before paying.

Because the proof is the production database, not a toy demo. A 57,900-point HyperspaceDB has been running as my personal knowledge base — real queries, real retrieval, real Lorentz coordinates. The 31k+ I quoted is a baseline; I rebuild and grow it. Your drop runs on the same pipeline.

Common case. The viz will show partial structure — some clear branches, some uniform clusters. The 3-page report calls this out: which subsets of your corpus benefit from hyperbolic, which don't. You decide whether to pursue. No automatic refund here, because you got real signal — the answer is just "mixed".

For Notion: a CSV mapping page_id → cluster_id → suggested_tag, plus a script that bulk-applies tags via the Notion API (you keep the API key). For Obsidian: a script that writes #tag annotations into the YAML frontmatter of each note, idempotently. Plus 10–20 suggested NEW tags that your current taxonomy is missing — based on dense regions in Lorentz space that don't correspond to any existing tag.

Me, Paulina Janowska, personally. The pipeline is my code. The 5 findings are written by me after looking at the viz. The landing page was drafted by my agent (Gniewisława AI / Hermes) — see disclosure below — but the deliverable is human work on real data.