Classification and indexing of content is a time-intensive process that end users loathe. As LLMs and AI/ML processing become more accessible, accurate, and precise, the “intelligent” part of “document processing” has become a feature that all ECM players are racing to implement into their solutions. This post demonstrates an AWS and Veladocs powered Intelligent Document Processing (IDP) flow in a common Insurance Claims scenario and shares how simple and elegant the integration is due to Veladocs cloud-native ECM architecture.

Insurance Claims Scenario

A typical insurance claim document lifecycle in a mature enterprise looks something like this:

1. Documents are imported and generated in a claims management system via automatic and manual means.
2. Those documents are ingested into a document case management system (DMS) where users may review document contents, document metadata, case previews, and more without having to download the documents.
3. Documents are archived for retention and legal purposes.

We are adding intelligence to the second step to demonstrate a simple intelligent document processing flow using Veladocs Search. In this scenario, the claims management system has stored and assigned case numbers and high-level data to the documents. We can enhance the DMS experience by classifying and enriching the documents by pulling additional document-specific metadata.

For a sample document below:

The claims management system already provides the following metadata:

• Claim Number
• Claimant Name

And likely a few other fields.

We’ve identified that the following fields would be helpful to have in our DMS:

• Document Category
• Document Date

We leverage the AWS services of Textract and Bedrock to classify and enrich our metadata, then use Veladocs APIs to ingest and index the documents.

Sample IDP flow

Upon upload to an S3 bucket, Python-based AWS Lambda functions will perform the following steps on object creation:

1. Extract fulltext document contents using Textract
2. Fetch required fields from Veladocs APIs
3. Request metadata values from Bedrock
4. Initiate document ingestion into Veladocs

Textract and Bedrock

Since Veladocs is a cloud-native solution, the steps to extract text and utilize AI to populate metadata values are exceedingly simple:

1. Request the text with a single API call to Textract.
2. Invoke Bedrock with a prompt specifying the metadata fields we are looking to retrieve (Document Category and Document Date) and including the fulltext of the document returned from Textract.

This naive AWS approach generates surprisingly good results in our testing with no additional upfront configuration or model fine-tuning.

After some refinement, here is our sample Bedrock prompt (using the amazon.titan-text-premier-v1:0 model):

Even as an R&D example, the level of precision and accuracy generated by this IDP flow may be good enough for many business scenarios. If results are not returning as expected, some simple steps may produce better results:

1. Refine the Bedrock prompt
2. Execute sequential prompts
3. Adding additional context to Bedrock prompts (source of document, other documents in the ingestion batch, etc.)

Veladocs APIs

1: What values need to be extracted from the documents?

The Veladocs platform handles three key steps for us in the IDP flow after the fulltext and metadata has been provided by Textract and Bedrock.

The Veladocs APIs store and serve representations of the object model and index fields for various document types segmented by “business context”. Because this document was uploaded into the Insurance Business Context, Veladocs APIs can identify that we need the aforementioned “Document Category” and “Document Date” fields.

2: How do the documents and metadata get indexed?

Veladocs provides indexing APIs that ensure standard ECM guarantees for content ingested into its systems, including the following:

• Enforcement and application of security
• Enforcement and application of the object model
• Set default metadata
• Search indexing and exposure to the interface and APIs

3: How do users find and review the content?

The Veladocs search interface provides a modern interface to search and view content. In the insurance claims case, users can execute a search and then quickly jump between results in the same case in their browser, saving significant time compared to navigating, downloading and viewing results from a claims management system. See results here of our sample claims with the classified and enriched content.

Check out the video below to see how we easily integrated these AI features into Veladocs as part of this effort:

Next Steps

Veladocs makes it easy to use the full power of your cloud platform provider’s capabilities for IDP. Here are a few of our thoughts on valuable next steps for IDP on Veladocs.

• Add value assistance lists to the LLM prompt for standardized classification and indexing
• Autodetect incorrectly located documents (e.g. a document mislabeled with the wrong claim number in the claims management system)
• Flag low confidence answers for manual intervention
• Cost/speed optimization
  • Extracting text and submitting only the first page of results
  • Using simple text/regex rules to classify/enrich documents without needing to send to LLM
• Further precision and accuracy testing on a variety of LLMs
• Create a custom ML model using AWS Comprehend
• Create a deep IDP flow (example: Intelligent document processing with Amazon Textract, Amazon Bedrock, and LangChain | Amazon Web Services )

Final Thoughts

Creating an IDP flow into Veladocs is simple with the combination of Veladocs APIs and the openness of the platform – if you can do it on AWS or Azure, you can do it in Veladocs without complex legacy APIs overcomplicating the design. Here at Docuvela, we combine our industry and solution expertise with the power and flexibility of our open cloud platform, Veladocs. Leave a comment with your IDP scenario below and reach out for us to help you set it up!