Versioned knowledge without whole-workspace publishing
Published 6 min read
Knowledge changes, but that should never erase who changed it, who approved it, or what came before. Curated Data versions each document independently, then makes an approved revision live immediately, without publishing an unrelated workspace bundle.
A company policy, metric definition, or source-system guide has its own lifecycle. Treating the entire knowledge base as one release couples unrelated work together: correcting one page would require rebuilding and publishing everything else. Per-document revisions keep the audit trail strong without adding that release bottleneck.
Approval publishes exactly one document
A change begins as a draft, moves into review, and becomes current when a reviewer approves it. That approval is the publication event. Authorized REST and MCP clients see the new revision on their next request; other documents remain exactly as they were.
Every revision keeps its identity
Revisions are numbered per document and record their author, change summary, review decision, reviewer, and timestamps. The document's stable path, for example policies/refunds.md, stays the same while its current approved revision advances. Prior revisions remain part of the history instead of being overwritten.
This gives reviewers a useful diff and gives operators a precise answer when someone asks what changed. It also keeps edits isolated: a pending update to the refund policy cannot hold back an approved correction to an analytics metric.
Agents only read current approved revisions
Retrieval never reads an author's open draft or a revision still waiting in the queue. Concept reads, search, and MCP tools all resolve through the same current-revision pointer and token classification checks. An agent therefore gets the latest reviewed content for each document, not the latest text someone happened to save.
What this model guarantees
- Approval makes one reviewed document live immediately.
- Unrelated documents are never republished as a side effect.
- Draft and in-review revisions remain unavailable to agents.
- Every document retains its numbered revision and approval history.
- A stable path gives integrations one identity across revisions.
Governance without release ceremony
Good governance should make the trustworthy action obvious, not make teams repeat it. Review is the meaningful gate: a named person examines a specific change and signs off. Publishing that same document at approval preserves the gate while removing a redundant whole-bundle workflow.
Next step
Put the approved answer behind every rep
Start free, load your packaging and pricing pages, and connect the assistant your reps already use.
Start freeKeep reading
- 7 min read
RAG retrieves. Review decides what your agent should trust.
Vector search can find relevant content without knowing which version is approved, current, or allowed. Here is how retrieval and curation fit together in a production agent.
- 8 min read
What belongs in an AI agent knowledge base?
A practical inclusion test, six high-value knowledge domains, a copyable concept template, and a first-week plan for building an approved source without importing every file.
- 7 min read
Why AI assistants guess, and how to give yours approved answers instead
When an assistant cannot reach the answer your company approved, it fills the gap from training data. Here is the mechanism that closes that gap: approved pages, traceable revisions, connected knowledge, and retrieval at answer time.
Put the approved answer behind your AI
Create a workspace, write and approve your first page, and connect an assistant over MCP or REST. Free to start, no credit card.
Not ready yet? Get future playbooks.