Custom Queries

Amazon Textract lets you customize the output of its pretrained Queries feature using adapters. You can use the Amazon Textract Console to create an adapter. This adapter can then be referenced when calling the AnalyzeDocument and StartDocumentAnalysis operations.

When you create an adapter using the console, you upload your own documents for the purposes of training the adapter and testing its performance. You also add queries to your documents and then annotate your documents by linking these queries to the correct response elements in your documents. Once you have created an adapter and annotated your documents, you can train the adapter, check its performance, and then use it when analyzing documents.

Adapters are modular components are added to the existing Amazon Textract deep learning model, extending its capabilities for the tasks it’s trained on. By fine-tuning a deep learning model with adapters, you can customize the output for document analysis tasks related to your specific use case.

To create and use an adapter, you must:

Uploading sample documents

To train the adapter, you must upload a set of sample documents representative of your use case. You can upload documents directly from your computer or an Amazon S3 bucket. For best results, provide as many documents for training as possible (up to a maximum of 2,500 pages training documents and 1000 test documents). Make sure that the documents represent all aspects of your use case. You must upload a minimum of five training and five testing documents.

Designating training and test sets

You must divide all of your documents into training and test sets. The training set is used to train the adapter. The adapter learns the patterns contained in these annotated documents. The test set is used to evaluate the adapter performance.

For more information on training and testing data, see Preparing training and testing datasets.

Annotating documents with queries and responses

When annotating your documents, you have two choices: You can auto-label your documents using the pretrained Queries feature and then edit the labels where needed. Alternatively, you can manually label responses for each of your document queries.

For more information on best practices for queries, see Best Practices for Queries.

Train the adapter

After you annotate the training data, you can initiate the training process for your adapter. Amazon Textract trains an adapter that's tailored to your documents. The adapter training takes 2-30 hours, depending on the size of the dataset and the AWS Region. When the training is complete, you can view the training status in the adapter details page. If the status is training failed, see Debugging training failures to debug the failure.

Evaluate the adapter

After each round of adapter training, review the performance metrics in the AWS Management Console to determine how close the adapter is to your desired level of performance. You can then further improve your adapter’s accuracy for your documents by uploading a new batch of training documents or by reviewing annotations for documents that have low accuracy scores. After you create an improved version of the adapter, you can use the AWS Console to delete any earlier adapter versions that you no longer need.

For more information on evaluation metrics, see Evaluating and improving your adapters.

Get the AdapterId

Once the adapter has been trained, you can get the unique ID for your adapter to use with the Amazon Textract document analysis API operations. Retrieve the AdapterId by using the ListAdapterVersions API operation, or by using the AWS Management Console.

Call the AnalyzeDocument API operation

To apply your custom adapter, provide its ID when calling the AnalyzeDocument or StartDocumentAnalysis API operations. This enhances predictions on your documents. When calling API operations, you can use up to one adapter per page.