@rnsk/brain
Workspace knowledge bases and KBrain — document RAG and memory injected server-side when enabled in chat.
Source: https://runstack.engineer/docs/rnsk-brain
Runstack KBrain combines document knowledge bases with lightweight workspace memory. It is internal to the server: the agent still uses only the four meta tools (searchTool, checkAuthentication, initiateConnection, executeTool). When KBrain is on, relevant excerpts and facts are injected into the system prompt before the model runs.
Configure vs use
| Step | Where |
|---|---|
| Configure | Workspace Settings → Knowledge Base — create collections, upload files, paste text, manage facts |
| Use in chat | Context popover → Runstack KBrain (persisted per workspace, like MCP server toggles) |
KBrain off = no embedding calls, no KB injection, no automatic memory writeback for that workspace.
Supported documents
.md, .txt, .pdf, .doc, .docx, .xls, .xlsx — parsed on the server with npm libraries, chunked for Cloudflare bge-small-en-v1.5 (512-token limit per chunk), embedded via Workers AI.
Requires Cloudflare Account ID + API token under LLM Providers (same credentials as Cloudflare chat models).
Agents
In Studio, when saving an agent you can attach knowledge base collections. If none are selected, chat uses all collections in the workspace when KBrain is enabled.
Memory facts
Facts can be added manually in Settings, extracted from user chat (simple patterns), or from successful tool JSON (stable keys like username, email, teamId). Facts are retrieved alongside document chunks when KBrain is on.
SDK
Programmatic @rnsk/brain npm package is not shipped yet; use the REST API under /api/knowledge/* from your workspace session or extend the server as needed.