> ## Documentation Index
> Fetch the complete documentation index at: https://docs.cortex-labs.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Apply an enrichment to a dataset

> Bind a selected published enrichment version, choose processing deliberately and inspect its results.

Use a published enrichment whose input kinds and output schema match your dataset. You need binding authority, sufficient clearance and an authorised execution principal.

## Select and bind a version

1. Open the dataset's **Analysis** page.
2. Select the enrichment definition and inspect its published version, expected inputs and evaluation evidence.
3. Create a dataset binding. Confirm the principal and displayed billing team.
4. Inspect context selection: the definition default, none, thread or neighbours, where supported.
5. Choose whether to run manually or enable automatic processing.

V2 uses dataset bindings. A workspace-and-dataset binding is refused. A dataset binding and its eligible results travel with that dataset to every workspace that reaches it.

Binding an LLM definition and enabling automatic processing are separate actions. A deterministic Format, Enrich or Aggregate bundle binding starts production work immediately. Use evaluated inputs and the intended dataset before binding one.

## Preview and run a bounded pass

For a controlled manual pass, preview the pending count before admission. Set a supported input limit or confirm the maximum pending count.

If the pending count exceeds your confirmation, no work is admitted. The count includes older pending work. A bounded pass processes at most its permitted inputs; a later pass handles the remainder.

A manual pass can run with automatic processing off. It leaves that setting off. A paused binding runs no pass, so unpause it deliberately when ready.

## Inspect results and omissions

Open representative results beside their exact source fragments. Check the schema, effective inputs, definition version and selected context.

Review pending, completed, failed and skipped inputs. A completed world-model extraction can still have claims dropped during evidence validation. Inspect those refusals before treating it as graph coverage.

An unchanged completed input reuses its result without another model call. A changed source, relevant neighbour/thread context or processing version can require new work.

## Upgrade, retry or pause

| Intended change | Action and check |
| - | - |
| Fix failed inputs | Run failed work after resolving its cause. Inspect the new outcomes. |
| Adopt another definition version | Change the selected version deliberately and inspect newly stale work. Existing runs retain their versions. |
| Recompute unchanged completed work | Use a forced rerun and record its purpose. It incurs separate computation. |
| Stop automatic work | Disable automatic processing or pause the binding. Ingestion can continue. |
| Resume after a pause | Check remaining source coverage. Expired inputs cannot be recovered by resuming. |

There is no action that clears all enrichment results or resets a binding's history. An expired result stays recorded as completed; unchanged work needs a forced rerun to recreate it.

Processing never lowers the dataset security level. Follow [privileged authority rules](/administration/people-and-access/privileged-authority) for authorised level changes.

For authoring and evaluation, use [Develop and evaluate an enrichment](/vantage/create-reusable-methods/develop-enrichment).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.