KOSHA
What KOSHA is
KOSHA is a database built for the way LLMs think.
Where a traditional database stores rows in tables, KOSHA stores facts in a small, fixed semantic shape — a 10-column pipe-delimited row called an ASHRU. Each fact carries the six Pāṇinian kāraka roles (who, what, with-what, to-whom, from-where, where) plus tense, negation, and free-form attributes. That structure happens to be ~80% smaller than the equivalent JSON when an LLM emits it, and ~80% cheaper on output tokens.
KOSHA bundles Google’s Gemma 4 so most queries never leave your laptop. When local resources run out, you can fall back to a cloud LLM with your own API key — KOSHA never sees the response.
What you get in the Community Edition
- A free Mac app (
.dmg) signed by SUMA Pro AI. - One workspace (NABHA) with eight default categories (TAARAs).
- Up to 8⁸ ≈ 16.7 million tree nodes per workspace.
- Local Gemma 4 (E2B by default; E4B on 16 GB+ Macs).
- Optional cloud fallback with your own OpenAI / Anthropic / Gemini key.
- Anonymous usage stats — opt-in, off by default.
What you don’t get (yet)
- Windows, Linux, mobile — Mac only for now.
- The retrieval ranking engine (K-WIL) — that ships in V2.
- Multi-device sync, hosted multi-tenant, or enterprise SSO — V2 / V3.
- Any data leaving your laptop without your explicit consent.
How the docs are organized
Following the Diátaxis framework — four kinds of documentation, each serving a different reader question:
- Quick start — install KOSHA and run your first query in five minutes.
- Concepts — the mental model. Start here if you want to understand the why.
- How-to — step-by-step solutions to specific tasks.
- Reference — the complete syntax of KoshaLang and the HTTP API.
If you can’t find what you need, open an issue and we’ll either add a doc or fix the bug.
Status
KOSHA is in private beta. Software may change without notice. The wire format (ASHRU) is locked at V1; the engine internals are still moving.