Summary
K4 keeps the product knowledge base current through agentic fetch, a ledger applier, retrieval over pgvector, and MCP bindings — live end to end, serving the real database read-only.

The knowledge tier, live

K4 is the quiet engine of the product: product knowledge that maintains itself. It is built in part — and live end to end:

  • Fetch — hourly runs land sha256-checked raw lines from named sources onto an append-only pad. An atomic publish (temp file + fsync + link, existing name refused) means a half-written line never lands.
  • Apply — a ledger applier applies KB SQL through the recorded ledger, idempotent by unique constraint, so every row traces to the run that fetched it.
  • Retrieve — PostgreSQL + pgvector with HNSW, trigram matching, and a local reranker.
  • Serve — the KB app serves the real database read-only, with MCP bindings for agents and a review board for humans.

Why it matters for adoption

A knowledge base that needs a curator dies at the first busy quarter. The design constraint here is not model capability — it is adoption cost: the extraction runs through SDK/MCP tooling and plugins that fit the tools the sales team already uses.

Still ahead

The promotion fence (what gets promoted from raw to published), the published-set runner, and the agent’s MCP-verbs door are the next three pieces of the governance plane.