Skip to content

Text Extraction with Azure AI Document Intelligence

Note

This feature is currently in Preview. We encourage you to try it out and provide us with feedback.

Text Extraction with Azure AI Document Intelligence creates semantic chunks, resulting in more accurate answers compared to the default fixed-length chunking.

When you upload a file with the .preset_ca suffix as a Knowledge Source, Cognigy automatically sends a request to the Azure AI Document Intelligence platform. This request prompts the platform to extract text from the file. Once the platform finishes processing the file, Cognigy retrieves the preprocessed result. This result is then converted to Markdown format and segmented into chunks for easier handling.

The chunk strategy powered by Azure AI Document Intelligence has the following advantages:

  • Improved quality of responses. Breaking down the text into blocks of meaning allows the virtual agent to more accurately understand the context and find the necessary information to formulate a higher-quality response.
  • Effective use of Markdown. The extracted text is presented in Markdown format in the Chunk editor. Markdown allows the virtual agent to recognize headers, tables, images, links, and differentiate them from a regular text. This approach adds an additional level of context, helping the virtual agent better understand the structure and organization of information in the text.
  • Improved ability to reference the source. Adding the page number of the source file to the metadata for each chunk can help track an information source in lengthy documents.

Supported Formats

Only files in PDF and DOCX are supported.

Availability

This feature is available in the following environments:

  • Trial.
  • Dedicated and shared SaaS. Only for a limited number of customers. If you want to use this feature, contact Cognigy technical support.
  • On-premises. To use this feature, set up Azure AI Document Intelligence connection and provide an API key.

On-Premises

Prerequisites

  • Cognigy.AI 4.71 is deployed.

Environment Variables

The following environment variables must be specified in the values.yaml file to enable the feature:

  • FEATURE_ENABLE_AZURE_DOCUMENT_INTELLIGENCE_ORG_WHITELIST:
    • This variable should be set to a comma-separated list of organization IDs for which the feature should be enabled. For example: FEATURE_ENABLE_AZURE_DOCUMENT_INTELLIGENCE_ORG_WHITELIST: "org1,org2,org3".
    • If the feature should be enabled for all organizations, the value can be set to "*".
  • AZURE_DOCUMENT_INTELLIGENCE_ENDPOINT. The endpoint URL for the Azure AI Document Intelligence service. To copy the URL, go to the Azure portal, on the left-side menu, select Resource Management > Key and Endpoint. Copy the endpoint URL from the Endpoint field. For more information, read the Azure documentation.
  • AZURE_DOCUMENT_INTELLIGENCE_APIKEY. Your API key for the Azure AI Document Intelligence service. To copy the key, go to the Azure portal, on the left-side menu, select Resource Management > Key and Endpoint. Copy the key from the Key field. For more information, read the Azure documentation.

How to Use

This strategy is applied when .preset_ca is appended to the file name of an uploaded file. For example, cognigy.preset_ca.pdf, where cognigy is the initial file name, .preset_ca is the preset to apply the text extraction, and .pdf is the file extension.

Metadata

In the Chunk Editor, each chunk has a metadata object that indicates the number of the source page where this information was located. The presence of page numbers in the metadata means that the text extraction feature works correctly. However, if there are no page numbers, it suggests an error, such as failing to access Azure AI Document Intelligence.

In the example below, the metadata object indicates that the chunk's text starts on source file page 1 and ends on page 2.

{
  "firstPage": "1",
  "lastPage": "2"
}

More Information