IVA OS · Persistent knowledge graph

Token-free answers,
straight from the graph.

The knowledge base below is an OKF (Open Knowledge Format) bundle — a directory of markdown files, one concept each, linked into a graph. Questions are answered by traversing that graph in your browser — resolve the entity, follow the edge, return the value. For facts in the graph, no language model is called and no tokens are spent. Beyond exact facts, a hybrid retrieval (graph + vector) and a model tier take over.

The stack. Knowledge format OKF · storage Git · retrieval hybrid RAG (graph + vector) · model Claude Fable 5 → Kimi K3 fallback. Answering escalates only as needed: exact graph fact (0 tokens) → vector retrieval over the KB (0 tokens) → a model grounded in what was retrieved (uses tokens; configure a key below). The graph is a curated seed bundle, so the split is deliberate — known facts stay token-free, the rest is the model's job.
⚙ Model settings — Claude Fable 5 (primary) → Kimi K3 (fallback)

The graph and vector retrieval always answer token-free. For questions they can't cover, the page calls a model — Claude Fable 5 first, falling over to Kimi K3 if it errors. Keys are stored only in this browser (localStorage), sent directly to each endpoint, and never committed or seen by AI Humane. Direct browser calls to these APIs are usually CORS-blocked — deploy the proxy and set each endpoint to your Worker URL.

Primary — Claude Fable 5

Fallback — Kimi K3 (open-weight)

Retrieval — neural embeddings (optional; default is token-free TF-IDF)

In the graph: