Live demos, recent updates, and a look at what the team is working on next

A lot of enterprise information still lives in documents that were made for people to read rather than for machines to process. PDFs, slide decks, scanned contracts, technical manuals and spreadsheets all carry information in their structure, not just in the words on the page. A person can usually tell which number belongs to which column, which caption belongs to an image, or which paragraphs sit under a particular heading. When a document is reduced to plain text, some of that context can easily get lost.

That matters more as companies try to use larger document collections for RAG, search, agents and other AI applications. A frontier model can work well with an individual report, but preparing thousands or millions of pages is a different problem. At that point, teams need a way to process documents consistently and turn them into something that can be reused across downstream systems. Roughly 90 percent of enterprise information is unstructured, so this is not a small part of the data problem.

Docling was developed to help with that part of the workflow. It started at IBM Research as an open-source document processing project and is designed to preserve things such as layout, tables, reading order, headings, hierarchy and metadata as a document is converted. The output can be written to structured formats such as Markdown, JSON and HTML, which makes it easier to use in retrieval, search and agent workflows.

OCR is part of the process when it is needed, particularly for scanned or image-based documents, but recognizing the text on a page is only one piece of the problem. The other part is understanding how the page is put together and preserving enough of that structure that the output still represents the original document.

For many teams, open-source Docling is a good fit. It can be run locally, configured for a particular workflow, and integrated into an application without relying on a managed service.

The requirements tend to change once document processing becomes part of a production system. There is more to manage around capacity, processing jobs, updates and the infrastructure running underneath it. That is where Docling for IBM watsonx comes in.

Docling for IBM watsonx provides the document processing capabilities as a managed service. Documents can be uploaded and inspected through the UI, or the service can be integrated into an application through its APIs. It also supports multi-document and batch conversion, including large collections stored in S3 or provided through web URLs.

The service has been available for a while now, and both the managed offering and the broader Docling project have continued to develop since launch. Some of that work is focused on the day-to-day realities of using document processing in production. Other changes are more visible in the product itself.

That is what we want to spend most of the upcoming webinar on.

We’ll walk through Docling for IBM watsonx using real documents and show what the workflow looks like from the initial input through to the structured output. We’ll use both the UI and APIs, look at how different kinds of documents are handled, and spend some time on the differences between running Docling yourself and using it as a managed service.

We’ll also show some of the capabilities that have been added since launch, discuss some of the connector work underway, and give a preview of a few areas the team is currently working on. Some of that work is still in development, so we’ll save the details for the session.

For anyone already using Docling OSS, the webinar should give a clearer view of what the managed service adds and where it may make sense to use one versus the other. For people who are newer to Docling, we’ll also cover enough of the basics to show how document structure is preserved and how the resulting output can be used in downstream AI applications.

The webinar will be on September 3rd at 11am Eastern.

Register for the webinar → https://ibm.webcasts.com/starthere.jsp?ei=1770721&tp_key=c4741dfc29