Put a collection on a schedule so its AI-written descriptions and READMEs are filled in for new assets and regenerated when an asset changes.
Keep collection fresh
Keep this collection fresh runs the Description and README agents on a collection on a schedule you choose. Each run fills in assets that have no AI text yet, and regenerates AI-written text on assets that have changed since it was written. Text a person wrote is never touched.
Prerequisites
Before you begin, make sure you have:
- Access to Context Agents Studio, which comes with the Admin or Governance Admin role. Anyone with access can keep a collection fresh.
- A collection to keep fresh. Both default and custom collections can be kept fresh.
- Flexi Credits available. Each refresh is charged like a manual run. See What refresh costs.
Turn on keep fresh
- In your Atlan workspace, click Context Agents Studio from the left sidebar.
- Open the collection you want to keep fresh.
- On the Keep this collection fresh card, turn the switch on.
- Under What to keep fresh, select Descriptions, READMEs, or both.
- Under When to refresh, choose how often to run and the timezone. The schedule starts as daily at midnight in your time zone.
- Click Turn on.
The card now reads On with what it keeps fresh and when. To change either, click Edit schedule, make your changes, and click Save. To stop, turn the switch off.
What happens on each run
Each run looks at the collection's 1,000 most popular assets and decides, asset by asset, what needs work:
| Asset | What the run does |
|---|---|
| No AI description or README yet | Generates one |
| AI-written text, and the asset changed since it was written | Regenerates it |
| AI-written text, nothing changed | Leaves it |
| Description or README written by a person, or crawled from your source | Leaves it, always |
When both agents are on, descriptions run first and READMEs after, so a README is written from the asset's latest description.
What counts as change
An asset counts as changed when any of these differ from when its text was written:
- Columns: a column added, removed, renamed, or given a new data type
- Definition: the SQL behind a view, materialized view, or dynamic table, a dbt model's SQL, or a query's text
- BI field formulas: for example a Power BI measure expression or a Tableau calculated field
- Linked glossary terms: a term linked or unlinked
- Linked knowledge files: a file linked or unlinked
Usage and popularity, owners, tags, certificates, and custom metadata don't count as changes. They move often, and regenerating on them can spend credits without changing what the text needs to say.
Each run checks for changes when it starts, reading the same metadata the agents read, so a change is picked up once it has been crawled into the metadata lakehouse.
Columns
When a table's description is regenerated, the run also writes every column that has no description or only an AI-written one. Columns with a description written by a person, or crawled from your source, are left as they are.
What refresh costs
Each run is charged the same way as Enrich Now: per asset and column it writes. Before it starts, the run works out what it writes and checks your workspace's credit balance.
If your workspace doesn't have enough credits for a refresh:
- The run doesn't start, and nothing is charged
- Keep fresh pauses itself for the collection, and the card reads Off · paused for credits
- A note on the card says when it paused, what that refresh needed, and what the workspace had, for example a refresh needed 88,170 credits and the workspace had 200
The note also compares that with your balance today. If the workspace now has enough credits, it says you can turn keep fresh back on. If it still doesn't, it says how many credits the refresh needs.
Keep fresh doesn't turn itself back on when credits are added: turn the switch back on when you're ready. If you'd rather leave it off, click Keep it off to clear the note. You can't turn keep fresh on while the workspace has no credits left.
Check last run
The card shows the last run and how it ended:
| Shows | What it means |
|---|---|
| Refreshing now · started 2 minutes ago | In progress |
| Last run 9 minutes ago | Finished and refreshed what had changed. The Activity tab shows what each agent wrote |
| Last run … · Nothing had changed | Every asset was already up to date, so nothing was written or charged |
| Last run … · Couldn't start: reason | The run didn't start. The reason says why, for example that the metadata lakehouse couldn't be read |
| Last run … · Enrichment failed | The run started but didn't finish its work |
| Last run … · Didn't finish | The scheduled run didn't complete |
Click View activity to see every refresh on the collection's Activity tab:
- Each agent run a refresh started appears as that agent's run, with what the refresh found, for example Description agent refreshed · Enriched 64 assets · 2 changed, 3 new
- Agents that didn't run appear as one Scheduled refresh row: nothing to update with how many assets were checked, paused when credits ran short, or couldn't start with the reason
- While a refresh is checking assets for changes, before any agent starts, a Scheduled refresh · checking for changes row shows at the top
Limits
- 1,000 assets per run: each run covers the collection's 1,000 most popular assets. See How many assets can be enriched per run?
- Replacing a knowledge file's contents under the same name isn't detected. Unlinking and relinking it is.
See also
- Enrich metadata on asset collections: Run an agent on a collection once, on demand
- FAQ - Metadata enrichment: Common questions about cost, limits, and overwrite behavior