Changing Context Policy does not re-embed. The collection profile has to match the vectors you already stored.

Last updated: 2026-11-05 Product area: Cookbook Page type: Tutorial

Index update workflow bound to an example collection — Indexing is a published workflow step. The screenshot values are an example profile, not a required model.

Index readiness

Move harbor-note.txt from extracted text to an indexed collection. Hands-on time is about 20 minutes. Embedding and index jobs are extra.

Purpose

You can say whether search is allowed to run. Level: editor. Role: workflow publish plus the processing permission your workspace requires. Mode: stored. Read Index a collection for the canvas clicks. This page is the readiness contract.

Prerequisites

Extraction has produced Smart Bites or the document has passed chunked. The embedding profile on the collection is the one already used for its vectors. Do not copy a screenshot’s provider, model, or dimensions onto a collection that was built with another profile.

Sample

One text file, one approved schema, zero expectation of a special index screen. Bundata does not have a separate Indexes page that you enable with a toggle.

Steps

On the collection, read status. ready means text is available. indexed means vectors were written. error means stop and read the run.
Open the index-update workflow, bind this collection, and match the existing embedding profile.
Save the draft, publish a release, then run that release. A saved graph is not an index.
Return to Catalog. Use the collection for search only when status is Indexed and the vector count is greater than zero.

Fields

Signal	Means search can run?
`processing` on the document	No. The file is still in the pipeline.
`ready`	Text only. Not vectors.
`indexed` and vector count > 0	Yes, for that collection.
Context Policy model or topK change	No rebuild. The next answer changes. The vectors stay.
Workspace default embedding change	Does not migrate this collection.

Example

There is no public “create index” JSON body to post. The runnable unit is the published workflow release. Record the release id and the run id.

Recovery

If the run errors before indexed, fix the profile mismatch or the missing provider key and run the published release again. Do not flip a non-retaining workspace to stored to force vectors. A provider key for Anthropic does not satisfy an OpenAI embedding profile.

Limits

No Haystack-style index object, no Enable button, and no promise that a policy edit re-embeds.

Check

Success: Catalog shows Indexed and a non-zero vector count for the example collection. Failure: status stays ready or error, and semantic search is not attempted.