# Overview

> The evaluation process for IXP model versions in the Extract node of UiPath Maestro™ Flow, covering performance evaluation and accuracy checks required before deploying a model to production.

Evaluating model performance is critical to ensuring the accuracy and reliability of the model before it is deployed to a production environment.

## The model evaluation process

**Prerequisite:** Annotate documents. You need ground truth for a representative set of documents to compare the predictions of the model against.

1. [Run an evaluation](running-an-evaluation.md) on a select set of documents.
2. Analyze model performance by reviewing the field-level and document-level results of the evaluation run. Both help you identify the factors that limit the accuracy of your model.
3. Act on your findings: try a different model configuration, iterate on the [field instructions](generating-and-reviewing-predictions.md#modifying-instructions) in your taxonomy, or correct annotations.
4. Rerun an evaluation, either on the same set of documents or on a specific subset you want to test.
5. Compare the performance of the two evaluation runs. When comparing model versions, we recommend using the same dataset.
6. Repeat this process until the model meets the performance targets it needs to be successful in production.

## Where to view model performance

The following parts of the product show the metrics and scores of your model:

* **Health score** - Shown on the Flow **Extract** node in two places: the properties pane and the **Definition** tab in the immersive takeover. It also shows which model version the score came from and which extraction model was used.
* **Eval runs history** - Lists every evaluation run for the node, with the F1 score and document error rate of each run. Select a run to see its details.
* **Field performance table** - Available in the **By field** view of a run. Shows one row per field or field group, with performance and error metrics calculated across the annotated documents.
* **Document performance table** - Available in the **By document** view of a run. Shows one row per document, with its F1 score, document error rate, total errors, total annotations, and annotated values.

For details about each metric, check [Evaluating model performance](evaluating-model-performance.md).

## Best practices

* Annotate at least 10 values for each field in the taxonomy, across as many unique documents as possible (ideally 10), to get a meaningful model and field score.