Short, dated notes on capabilities that just landed in the platform.
Search covers titles, summaries, tags and full article text.
KDCube now prepares every app under its own supervised lifecycle. A slow checkout, hook, index, or UI build makes that app temporarily unavailable without making an otherwise healthy processor look dead.
A vector is a pure function of query, model and dimension — so every repeated search re-pays for an answer the system already has. The obvious fix is to remember it on the index object, and in a runtime where turns hop between workers it catches the one repeat that rarely happens. What the shared cache needs instead: a key that carries the model, a value ...
Switch a tool off and every trace of it leaves the agent's instruction — teaching blocks, protocol channels, catalog sections. One authority decides: the effective tool roster.
How a KDCube app verifies Telegram webhooks, admits updates to its event lane, runs its configured agent, and delivers results back.
The KDCube ReAct agent's full instruction body is ~24,800 tokens of battle-proven prose. It now has two distilled siblings — a moderate set at 40% and an extra-lite set at 30% — built by one discipline: enumerate every hard signal, then cut only the prose around it. The signal table is executable: it lives in tests that fail if a distillation loses a sing...
Point a KDCube agent at a model on your own machine — served by Ollama — and let a user pick it for a single conversation. A small models gateway translates the existing provider: custom path; a reserved descriptor property turns it on.
An agent's instructions compose from blocks, so it can offer named profiles a user picks per conversation — including a distilled set for local models.
The platform renders catalog pages; until now an application could recolor exactly three things. This entry records the presentation contract that ends that: every visual decision in the renderer reads a --kdcpub-* design token, and an application overrides tokens or loads its own stylesheets — declared in its config, per alias and per fold. The platform ...
KDCube counts LLM calls, embeddings, and web searches as usage — recorded, priced, and settled against a user's budget. Your application's own paid work can join them. A decorator, four small extractors, and one usage object turn a metered API call into first-class usage the economics model can charge.
Bring your own agent — your graph, your framework, your control loop. KDCube hands it user events through the event bus: ordered and one turn at a time per conversation. A run-to-completion loop receives the next accepted start batch as its next turn; a loop that can absorb input mid-flight may fold new work into the live turn.