A living knowledge grid that structures your documents, retrieves facts, reasons an answer — and verifies whether that answer is true, in zero tokens.
문서를 3D 온톨로지 그래프로 구성하고, 질문하면 그래프 사실을 근거로 LLM이 답하며, 그 답을 ZTC가 0토큰으로 판정합니다.
ONGRID in one line: an open ontology graph + GraphRAG engine that renders knowledge as a glowing 3D universe, answers from graph facts, and verifies every answer with a zero-token judge (ZTC).
Most GraphRAG tools retrieve and generate. ONGRID adds the missing step the whole industry is scared of: decision — telling you whether the answer is actually true.
| Stage | What happens | Tech | |
|---|---|---|---|
| O·N·G | Ontology-Native Graph | Documents → typed entities + relations → one 3D graph | three.js · 3d-force-graph |
| R | Retrieval | Pull the facts relevant to your question | GraphRAG |
| I | Inference | An LLM answers only from those facts | LLM API |
| D | Decision | ZTC-Judge scores if the answer is trustworthy — 0 generated tokens | FINAL-Bench ZTC-Judge |
🪐 Structure → 🔎 Retrieve → 🧠 Reason → 🛡️ Decide
Roadmap: a self-evolving grid — it grows and refines itself as you feed it (RSI).
The default universe is built from 227 real press articles (models, benchmarks, media, topics). Click an example query and watch the graph focus, the grounded answer, and the ZTC verdict badge appear.
| Ask & get a grounded answer, verified by ZTC | The whole knowledge universe |
|---|---|
 |
|
Ask "Which model tops S1MB?" → the graph focuses, the LLM answers only from retrieved facts, and the answer is verified in zero tokens.
- 🌐 Trilingual — 한국어 · English · 中文, auto-detected, one-click switch.
- 🪐 3D ontology graph — typed nodes, curved glowing edges, clustering, search, click-to-focus.
- 🧠 Grounded answers — the LLM answers only from retrieved graph facts, with evidence shown.
- 🛡️ Zero-token verification — every answer judged by ZTC-Judge in one forward pass.
- 📁 Bring your own docs — upload
.md / .txt / .json / .csv; parsed in your browser, nothing uploaded. - 🧱 Open & on-prem — Apache-2.0, runs anywhere Docker runs.
git clone https://github.com/openaiteams/ongrid.git
cd ongrid
export ANTHROPIC_API_KEY=sk-ant-... # never commit keys
docker build -t ongrid . && docker run -p 7860:7860 -e ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY ongrid
# open http://localhost:7860| Env var | Default | Meaning |
|---|---|---|
ANTHROPIC_API_KEY |
— | LLM key for grounded answers |
ZTC_REPO |
FINAL-Bench/ZTC-Judge-4B |
ZTC-Judge model (-9B/-27B on GPU) |
KNOU_MODEL |
claude-haiku-4-5-20251001 |
answer model id |
What is an ontology graph?
Knowledge represented as typed entities and typed relations — concepts and exactly how they connect. ONGRID builds one automatically from documents and renders it in 3D.What is zero-token verification (ZTC)?
ZTC judges whether an answer is trustworthy in a single forward pass, generating no tokens. ONGRID grades every answer so you see a trust verdict, not just text.How is ONGRID different from other GraphRAG tools?
It adds a decision layer: after retrieve-and-answer, ZTC-Judge verifies the answer. It is also fully 3D, trilingual, and runs your docs in the browser.- 🤗 Live demo: https://huggingface.co/spaces/FINAL-Bench/ONGRID
- 🛡️ ZTC-Judge: https://huggingface.co/FINAL-Bench/ZTC-Judge-9B
- 🏆 S1MB leaderboard (ZTC #1 of 102): https://huggingface.co/spaces/hotchpotch/S1MB-leaderboard
- 🌐 VIDRAFT: https://vidraft.net
- 🌐 Website — https://vidraft.net
- 🤗 Hugging Face — FINAL-Bench · VIDraft
- 🧵 Threads — @vidraft_lab
- 𝕏 X — @VIDRAFT_ai
- ✉️ Email —
vidraft@mail.vidraft.net
Issues and PRs welcome — new extractors, languages, visual themes, ZTC integrations.
Apache-2.0 · Built by VIDRAFT · Beyond AI, Toward AGI.