OriginOriginChainDB / Guides
Saved embedders and embedding runs
Generate vectors from selected table columns, schedule updates and search with plain text using a saved embedder.
Search with a question
After a table has been filled by an embedding run or migration with a saved embedder, search its vectors with plain text from the console or API. OriginChainDB embeds the question with the same saved embedder and returns nearby rows.
This requires a prepared table and a saved embedder; supplying arbitrary vectors alone does not configure text search. See the 19 October release update.
Save and test an embedder
In organization Settings → Embedders, save the endpoint URL, key and model, then test the connection before using it. Use a public HTTPS endpoint; local and private-network URLs are not accepted. The key is encrypted, and only its last characters are displayed later. Changing the URL requires entering the key again.
Selected text is sent to the saved endpoint. Review the provider’s data handling and charges before starting a run. Keep credentials out of the text columns you select.
Embed a table already on an instance
- Open the dedicated instance’s Platform → Embeddings page and start a new run.
- Choose a table keyed by one column, and select the columns to embed in the order they should be joined.
- Choose a tested saved embedder. Optionally add row conditions; all supplied conditions must match.
- Choose which rows to process and set the required maximum number of rows to embed.
- Start the run and follow its state, rows read, rows embedded and any failure reason.
| Mode | Behavior |
|---|---|
| Rows without a vector | The default. Run again later to process rows that still have no vector. |
| New and changed rows | Also regenerates vectors when their source text changed. |
| Every row | Regenerates vectors for all selected rows, subject to the run limit. |
New and changed rows and Every row replace a stored vector together with its metadata. Other metadata under the same row ID is not retained. Treat metadata keys beginning with oc_ as reserved.
Permissions and data visibility
A run reads the table as the member who started it, applying that member’s row policies and column masks. A masked key column or text column to embed is refused. Members who can operate the instance can start, cancel or restart runs; other members can view them.
Cancel or continue a run
Use Cancel to stop a run. For a failed or cancelled run, Start again continues from its saved progress and keeps the count toward its row limit. Read the failure reason before restarting; an invalid endpoint, missing permission or unsuitable table needs attention first.
Keep embeddings current
Set up an embedding schedule to process new rows automatically and, optionally, rows whose source text has changed. Without a schedule, start an explicit run when you want vectors refreshed. A row edit alone does not start an unscheduled job.
Embed during migration
The Embed rows option in Migrate uses a tested saved embedder and a table keyed by one column. Row conditions choose what gets embedded; they do not exclude those other rows from the copy. Review the selected text, conditions and maximum row count before approving.
For common questions, see the migration and embeddings FAQ.