Version Snapshots and Content Organization
Create and preview knowledge-base snapshots, organize active document versions, and select a content manifest for marketplace review.
A knowledge-base snapshot records the document list and each document's active content version at creation time. Its primary purpose is to provide a reviewable content selection that a marketplace submission can target.
A snapshot is not a complete backup or a copy of the Milvus index. The current interface cannot roll a knowledge base back from a snapshot, and it does not provide snapshot deletion.
What a snapshot records
When a snapshot is created, the platform collects documents that are processing or indexed and records:
- Document ID and title
- The active Content Version ID, when one exists
- Version label
- Source type, such as original or a Data Foundry de-identified version
- Creator and creation time
It does not duplicate the complete vector index, and it does not prevent the source knowledge base from continuing to add, edit, or remove documents.
Organize active content first
Before creating a snapshot:
- Wait for the documents you need to finish parsing and indexing.
- Open each document's Chunks page to remove irrelevant content and correct inaccurate chunks.
- If Data Foundry is used, confirm that the intended de-identified content version is active.
- Test representative questions with knowledge-base Search.
The snapshot record itself is not rewritten when an active version changes. Entries with a Content Version ID continue to point to that version. An entry without one stores a reference to the original document rather than a full content copy, so a snapshot should not be treated as an immutable file backup.
Create and preview a snapshot
Open the knowledge base and go to Settings → Snapshots.
Select Manage snapshots, then Create snapshot. The label is generated from its creation time.
Select Preview beside a snapshot to review its file count, document names, version labels, and content previews.
Only the creator or a collaborator with edit permission can create and manage snapshots. Members with view access can read snapshot information for an accessible knowledge base but cannot change its submission state.
Submit a snapshot for marketplace review
A marketplace submission must target one snapshot:
Go to Settings → Publishing and set the knowledge base to public.
Complete the marketplace description, price, and the icon and cover fields required by the interface. A listing fee can apply according to platform settings.
Select a snapshot and submit after the review checklist is complete.
Submitting a different snapshot supersedes an existing pending submission. Editing knowledge-base content while a submission is pending also marks that submission as Superseded. Organize the updated content, create a new snapshot, and submit it again.
Submission states
| State | Meaning |
|---|---|
| Pending review | Submitted and waiting for an administrator |
| Approved | The snapshot is approved as the marketplace version |
| Rejected | Review failed; an available reason is shown on the Publishing page |
| Withdrawn | The pending submission or current marketplace publishing flow was withdrawn |
| Superseded | Content changed during review, or another submission replaced it |
Approved and newly pending snapshots are tracked separately. If a new version is rejected, that rejection does not clear an already approved snapshot.
Snapshots versus versions
| Item | Document or chunk version | Knowledge-base snapshot |
|---|---|---|
| Scope | One document content version or one chunk | The knowledge base's document and active-version list |
| Purpose | Edit, compare, and restore individual content | Preview and fix the target of marketplace review |
| Direct restore | An individual chunk version can be restored | The whole knowledge base cannot currently be rolled back from a snapshot |
| Creation | Saved by content-editing workflows | Created manually by a user |
Related documentation
Knowledge Base Overview
Data Foundry
Chunking Strategies
Knowledge Base Permissions
Data Foundry De-identification
Run de-identification jobs on knowledge-base chunks, monitor processing, and manage the resulting versions.
Querying Large Databases
Handing a database to an assistant — from connecting it, through letting the system read it, to the few sentences you write that make it answer accurately. Every step says why it exists, and whether what you wrote actually reaches the assistant.