Compare catalogs

There is no Compare Catalog button. Two collections are compared by a manual API call. Chat A / Chat B does not do this. The Workspace Graph’s similarity edges do not do this. Eval Studio does not do this.

Request

POST /api/w/:workspaceId/catalog-parity/compare

The caller needs a session and collection:read on that workspace. Both ids must be collections in that same workspace, and they must differ. A foreign or missing id is not found. The body does not select a workspace.

{
  "baselineCollectionId": "00000000-0000-4000-8000-000000000001",
  "candidateCollectionId": "00000000-0000-4000-8000-000000000002"
}

These ids are placeholders. They are not a Bundata API key, and there is no public SDK or download button for the report.

Report

The report has five dimensions: schema, extraction, smart_bites, vector, and drift_summary. Each is pass, warn, or fail. The overall status is the worst of those. Snapshot warnings mean the read of a collection was incomplete. A missing embedding profile or missing role metadata is a warning or a fail, not a silent match.

Schema compares the contract. Extraction compares extracted values. Smart bites compares counts and sizes. Vector compares index coverage. Drift summary names the first failing dimension as the likely cause. Suggested actions are advice. Nothing is repaired automatically.

A percentage-point gap is a difference of percents. A relative gap is a fraction of the baseline. Missing data is not zero. A pass does not mean the collections say the same thing, and it does not approve a release.

Evidence in this package

No isolated comparison was run for this delivery. The example above is the request shape only.