Large language models (LLMs) such as OpenAI’s GPT-4 are the building blocks for an increasing number of AI applications. But some enterprises have been reluctant to adopt them, owing to their inability to access first-party and proprietary data.
It’s not an easy problem to solve, necessarily — considering that sort of data tends to sit behind firewalls and comes in formats that can’t be tapped by LLMs. But a relatively new startup, Unstructured.io, is trying to remove the roadblocks with a platform that extracts and stages enterprise data in a way that LLMs can understand and leverage.
Brian Raymond, Matt Robinson and Crag Wolfe co-founded Unstructured in 2022 after working together at Primer AI, which was focused on building and deploying natural language processing (NLP) solutions for business customers.
“While at Primer, time and again, we encountered a bottleneck ingesting and pre-processing raw customer files containing NLP data (e.g., PDFs, emails, PPTX, XML, etc.) and transforming it into a clean, curated file that’s ready for a machine learning model or pipeline,” Raymond, who serves as Unstructured’s CEO, told TechCrunch in an email interview. “None of the data integration or intelligent document processing companies were helping to solve this problem, so we decided to form a company and tackle it head-on.”
Indeed, data processing and prep tends to be a time-consuming step of any AI development workflow. According to one survey, data scientists spend close to 80% of their time preparing and managing data for analysis. As a result, most of the data companies produce — about two-thirds — goes unused, per another poll.
“Organizations generate vast amounts of unstructured data on a daily basis, which when combined with LLMs can supercharge productivity. The problem is that this data is scattered,” Raymond continued. “The dirty secret in the NLP community is that data scientists today still must build artisanal, one-off data connectors and pre-processing pipelines completely manually. Unstructured [delivers] a comprehensive solution for connecting, transforming and staging natural language data for LLMs.”
Unstructured provides a number of tools to help clean up and transform enterprise data for LLM ingestion, including tools that remove ads and other unwanted objects from web pages, concatenate text, perform optical character recognition on scanned pages and more. The company develops processing pipelines for specific types of PDFs; HTML and Word documents, including for SEC filings; and — of all things — U.S. Army Officer evaluation reports.
To handle documents, Unstructured trained its own “file transformation” NLP model from scratch and assembled a collection of other models to extract text and around 20 discrete elements (e.g., titles, headers and footers) from raw files. Various connectors — about 15 in total — draw in documents from existing data sources, like customer relationship management software.
“Behind the scenes, we’re using a variety of different technologies to abstract away complexity,” Raymond said. “For example, for old PDFs and images, we’re using computer vision models. And for other file types, we’re using clever combinations of NLP models, Python scripts and regular expressions.”
Downstream, Unstructured integrates with providers like LangChain, a framework for creating LLM apps, and vector databases such as Weaviate and MongoDB’s Atlas Vector Search.
Previously, Unstructured’s sole product was an open source suite of these data processing tools. Raymond claims that it’s been downloaded around 700,000 times and used by over 100 companies. But to cover development costs — and placate its investors, no doubt — the company’s launching a commercial API that’ll transform data in 25 different file formats, including PowerPoints and JPGs.
“We’ve been working with government agencies and have several million in revenue in just a very short period. . . . Since our focus is on AI, we’re focused on a sector of the market that’s not affected by the broader economic slowdown,” Raymond said.
Unstructured has unusually close ties to defense agencies, perhaps a product of Raymond’s background. Prior to Primer, he was an active member of the U.S. intelligence community, serving in the Middle East and then in the White House during the Obama administration before a stint at the CIA.
Unstructured was awarded small business contracts by the U.S. Air Force and U.S. Space Force and partnered with U.S. Special Operations Command (SOCOM) to deploy an LLM “in conjunction with mission-relevant data.” Moreover, Unstructured’s board includes Michael Groen, a former general and director of the Pentagon’s Joint Artificial Intelligence Center, and Mike Brown, who previously led the Department of Defense’s Defense Innovation Unit.
The defense angle — a reliable early revenue source — might’ve been the deciding factor in Unstructured’s recent financing. Today, the company announced that it raised $25 million across a Series A and previously undisclosed seed funding round. Madrona led the Series A with participation from Bain Capital Ventures, which led the seed, and M12 Ventures, Mango Capital, MongoDB Ventures and Shield Capital, as well as several angel investors.