Knowledge Understanding & Summarization
Make a small model answer like a big one.
Kus is an external knowledge API. When a small or local model reaches the edge of what it knows, it calls Kus — which answers from a self-growing knowledge base, or agentically searches the web, learns, and answers with citations.
How it works
Your model hits its limit
A small or local model gets a question about tech, science or maths it can't answer well on its own.
It calls Kus as a tool
One endpoint, one API key. Kus classifies the question and fans it out across its knowledge base and the live web.
It answers, with sources
You get a grounded answer plus citations you can check — and whatever Kus learned stays in the knowledge base for next time.
Under the hood
Retrieval quality is the whole product, so it is worth being specific about how it is built.
Hierarchical parallel retrieval
A master router fans a query out to three child RAGs and nineteen async source workers, all concurrent — so breadth costs latency once, not once per source.
Two-stage ranking
Hybrid retrieval (BM25 + dense vectors, fused with RRF) narrows the field, then a FlashRank cross-encoder reranks what survives. It runs on CPU, so it adds accuracy without adding cost.
Semantic answer cache
A question close enough to one already answered returns in about 150ms with no model call at all.
Your documents stay yours
A shared public knowledge base plus private per-tenant documents, isolated by API key. Nothing you upload leaks into anyone else's answers.
Quality gate on what it learns
Auto-grown knowledge is reranked before it is kept, and weaker chunks are given shorter lifetimes — so the knowledge base doesn't entrench its own mistakes.
No single search vendor
Web search sits behind a provider seam — DuckDuckGo by default, SearXNG or Brave if you prefer. Both defaults are free and need no API key.