Document AI: Industrialize Your Document Processing Workflows with Snowflake Features

What is Document AI? 

Document AI is the flagship component of Snowflake’s Cortex AI family, enabling the industrialization of intelligent document processing through the Arctic-TILT LLM model developed by Snowflake.
Its scope is not limited to text paragraphs; it can also extract graphical content such as logos, signatures, or checkboxes. Integrated into the Cortex AI suite, Document AI enables users to leverage extracted information as context for the LLM model, providing targeted and precise responses.
The objective of Document AI is to democratize intelligent document processing and offer various support options to leverage this information, such as Streamlit and SQL analytics.

What are the components of Document AI?

Model Build UI:

The Model Build UI component of Snowflake’s Document AI, accessible through Snowsight, enables the creation and management of data extraction models specific to a document type, such as invoices.

Each model includes the model itself, the data values to extract, and the documents used to test and train the model.

The interface is divided into several tabs:

• “Build Details” for model information,
• “Documents” for the list of uploaded documents,
• “Values” for the data values to extract.

This structure facilitates the management and optimization of models for efficient data extraction.

Arctic-TILT Model:

The Arctic-TILT model is an LLM model that uses a unique transformer architecture to understand and extract data from documents by combining multiple modalities, such as text, images, and layouts.

It offers a holistic understanding of content and its context, with state-of-the-art performance on benchmarks like DocVQA, where it outperforms models like GPT-4 despite having fewer parameters.

Its extended context window of 375,000 tokens enables it to capture the complete context of multimodal content, while its inference efficiency allows it to handle varied document volumes while maintaining high accuracy.

Adaptable to various applications and industries, Arctic-TILT requires no prior knowledge of documents and can be easily fine-tuned to meet specific user needs.

!PREDICT Function:

The “!PREDICT” method of Snowflake’s Document AI enables the extraction of information from documents by executing an extraction query in worksheets.

Once the Document AI model is published or trained, users can execute this query to extract specific data from documents stored in internal or external stages. This method relies on a SQL query based on “!PREDICT”, accessible through the Snowsight interface.

To use this functionality, users must have the appropriate roles and privileges, and documents must be properly stored.

Additionally, Document AI enables the creation of automated processing pipelines to continuously extract information from new documents added to a stage, using streams and tasks.

This approach facilitates the integration of extracted data into operational workflows, thereby improving the efficiency and automation of document processes.

Business Process Optimization:  

In the context of business operations, companies often face challenges when they need to verify generated documents against their reference system.

Traditional manual processes, involving visual comparison of documents, are time-consuming and error-prone. To overcome these obstacles, some companies initially collaborated with vendors to develop platforms, but this required months of development and years of fine-tuning. By adopting an in-house Document AI solution, companies can reduce their dependence on third parties, decrease document processing costs, and gain flexibility.

The solution processes various document types with an easy-to-maintain architecture. This approach enables the automation and streamlining of document processing, thereby improving operational efficiency and data accuracy.

Register for our upcoming data breakfasts

Pre-register and receive the invitation as soon as a new session is scheduled.