Understanding validation on extractions and extraction performance

Note: The Extractions Validation page is in public preview.

The Validation page shows an overview of extractions performance, and helps you drill down into the individual performance of each extraction.

You can access the Validation page from the Extractions tab within Validation, and is only available if you have extraction fields defined on your dataset.​

Performance overview

The default page of the Extractions Validation page is the ‘All’ overview page, which provides the following summary stats on the overall performance of extractions in the dataset:

• Extractions Mean F1 score
• Extractions Mean Precision
• Extractions Mean Recall

Figure 1. Extractions Validation page

Note: A low Extraction Mean F1 score is a sign of poor extraction performance​.

Individual extractions

Click into individual labels to see the performance of individual extractions, i.e. the label and its associated extraction fields.

For each extraction, you can see the following values:

• F1 score
• Precision
• Recall

For all of the extraction fields for the label, you can see the following values:

• Average F1 score
• Average precision
• Average recall

For the individual extraction fields, you can see the following values:

• F1 score
• Precision
• Recall

Figure 2. Individual extraction validation

Confidence levels and LLMs

How the confidence levels work varies depending on the underlying LLM model that you use.

If you use the Preview LLM

The Preview LLM does not have confidence levels on its predictions. The Preview LLM returns whether a label or field is a prediction (Yes = 1), or not (No = 0).

As a result, there is no concept of different confidence thresholds.

If you use the CommPath LLM

If you use the CommPath LLM, the model uses its Validation capabilities to predict which labels to apply to a communication. The model assigns each prediction a confidence score (%). This shows you how confident is the model that the label applies.

Automating with Generative Extraction

Note: In case you have previously built automations with entities: If you are planning on building automations using Generative Extraction, the API endpoint and activities required to build this out are different than what you have previously used.​

This section describes the outputs of the get stream results activity. Check the Communications Mining dispatcher framework page for more details.​

To automate with Generative extraction, it is important to understand the contents of the outputs of your extractions.​

Occurrence confidence: Refers to how confident the model is around the number of instances a request might occur on a message (i.e.- how many times an extraction might occur). ​

As an example: To process a statement of accounts into a downstream system, you always need an Account ID, PO number, the payment amount, and the due date.

Check below the occurrence confidence example. It shows how the model can confidently identify that there are 2 potential occurrences where you need to facilitate this downstream process​.

Extraction confidence

Extraction confidence is the model's confidence about its predictions. This includes how accurate it thinks it was in predicting a label's instance and its related fields. It also includes the model's confidence in correctly predicting if a field is missing.

Consider the same example as before. To process a statement of accounts into a downstream system, you always need an Account ID, PO number, the payment amount, and the due date.

However this time, the PO number is not present on the message, or the due date (only the start date).

The extraction confidence from this example is the model's confidence about identifying if the values for each field associated with the label are present. It also includes the model's confidence in correctly predicting if a field is missing.

In this case here, you don’t have all the fields you need, to be able to fully extract all the required fields.

Payload example

Check below an example output of what the get stream response activity returns.

Stream refers to the threshold you set in Communications Mining, and if the message meets this threshold.

Instead of filtering out predictions based on thresholds, this route returns which prediction confidence met the thresholds.

In other words, if your thresholds were met, stream is returned. If not, then this value is empty.​

Note: This is a sample snippet to explain the different components, not the full output of a Generative Extraction response. ​

Additionally, where there are multiple extractions, it is conditioned on the extractions before it.​

For labels without extraction fields, the occurrence confidence is equivalent to the label confidence that you can see in the UI.​

Note: If the model has failed to successfully extract all of the fields in a message because there are too many fields, it will return an extraction in the stream response that has an occurrence confidence AND an extraction confidence with values of 0.