Unstructured, Offering Tools to Prepare Enterprise Data for LLM, Raises $25M | TechCrunch

Image credits: bulldog_invincible / Getty Images

Large language models (LLMs) such as OpenAIs GPT-4 are the building blocks of a growing number of AI applications. But some companies have been reluctant to adopt them, due to their inability to access first-party and proprietary data.

This is not an easy problem to solve, necessarily considering that this type of data tends to get behind firewalls and arrives in formats that cannot be exploited by LLMs. But a relatively new startup, Unstructured.io, is trying to remove the hurdles with a platform that mines and stages business data in a way that LLMs can understand and exploit.

Brian Raymond, Matt Robinson and Crag Wolfe co-founded Unstructured in 2022 after working together at Primer AI, which was responsible for building and implementing natural language processing (NLP) solutions for enterprise clients.

While at Primer, time and again, we’ve encountered a bottleneck in ingesting and preprocessing raw customer files containing NLP data (e.g., PDF, email, PPTX, XML, etc.) and turning it into a clean, curated file ready for a model or machine learning pipeline, Raymond, who serves as the CEO of Unstructureds, 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 set up a company and tackle it head on.

Indeed, data processing and preparation tends to be a time consuming stage in any AI development workflow. According to one survey, data scientists spend nearly 80% of their time preparing and managing data for analysis. As a result, about two-thirds of the data that companies produce is not used, according to another survey.

Organizations generate large amounts of unstructured data on a daily basis, which when combined with LLMs can boost productivity. The problem is that this data is scattered, Raymond continued. The dirty secret in the NLP community is that data scientists today still have to build one-off, artisanal data connectors and preprocessing pipelines completely manually. Unstructured [delivers] a complete solution for connecting, transforming and staging natural language data for LLMs.

Unstructured provides a number of tools to help cleanse and transform enterprise data for LLM import, 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 all things related to US Army officer evaluation reports.

To handle documents, Unstructured trained its file transformation NLP model from scratch and assembled a collection of other models to extract text and about 20 discrete elements (e.g. titles, headers, and footers) from raw files. Various connectors, about 15 in total, pull documents from existing data sources, such as customer relationship management software.

Behind the scenes, we’re using a variety of different technologies to abstract out complexity, Raymond said. For example, for old PDFs and images, we used computer vision models. And for other file types, we used clever combinations of NLP patterns, Python scripts, and regular expressions.

Downstream, Unstructured integrates with providers such as LangChain, a framework for building LLM apps, and vector databases like Weaviate and MongoDBs Atlas Vector Search.

Previously, Unstructured’s only product was an open source suite of these data processing tools. Raymond says it has been downloaded approximately 700,000 times and used by over 100 companies. But to cover development costs and appease its investors, the company is no doubt rolling out a commercial API that will transform data into 25 different file formats, including PowerPoint and JPG.

We have worked with government agencies and made several million in a very short period. . . . Because we focus on AI, we’ve focused on a sector of the market that isn’t impacted by the broader economic slowdown, Raymond said.

Unstructured has unusually close ties to defense agencies, possibly a product of Raymond’s background. Prior to Primer, he was an active member of the US 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 US Air Force and US Space Force and worked with the US Special Operations Command (SOCOM) to implement an LLM along with mission-relevant data. In addition, Unstructured’s board includes Michael Groen, a former general and director of the Pentagon’s Joint Artificial Intelligence Center, and Mike Brown, who formerly led the Department of Defense’s Defense Innovation Unit.

The defense angle, a reliable source of upfront revenue, may have been the deciding factor in Unstructured’s recent funding. Today, the company announced it has raised $25 million through a previously undisclosed Series A and seed funding round. Madrona led 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.