Contextual AI Inc., a startup that helps enterprises build retrieval-augmented generation artificial intelligence applications, has closed a $80 million Series A round to support its commercialization efforts.

The company disclosed in its announcement of the raise on Thursday that Greycroft was the lead investor. The venture capital firm was joined by more than a half-dozen other backers, including the startup investment arms of Nvidia Corp., Snowflake Inc. and HSBC Holdings plc. Jeff Bezos’ Bezos Expeditions participated as well. 

Contextual AI is led by Chief Executive Officer Douwe Kiela (pictured, center), who was part of the Meta Platforms Inc. research team that invented RAG in 2020. Before the discovery, large language models could only answer questions using the information in their training datasets, which limited their usefulness. RAG is a machine learning technique that allows LLMs to draw on information from external data sources.

Contextual AI’s flagship product is a software platform that eases the task of building RAG applications. According to the company, the platform implements an improved version of the technology it dubs RAG 2.0. Contextual AI says that its version of the data retrieval method enables AI applications to answer user queries more accurately than the original version. 

A RAG system comprises three main components. There’s an LLM tasked with answering user questions, as well as a software module called a retriever that supplies the LLM with information from external data sources. The AI model incorporates that external data into the responses it generates. 

The third core component necessary to build a RAG application is an embedding model. This is a specialized neural network tasked with turning raw data, such as business documents, into mathematical structures that LLMs can process more easily.

Developers often create RAG applications by cobbling together an LLM, retriever and embedding model from different providers. Software components developed by different companies frequently aren’t optimized to work with one another out of the box. As a result, linking them together into a functioning application can be a highly complicated process. 

According to Contextual AI, its RAG 2.0 technology simplifies the workflow. It makes it possible to provide the LLM, retriever and embedding model that make up a RAG system as a single, pre-optimized bundle. Customers are thereby spared the hassle of manually cobbling together software components from different providers, which saves time and lowers development costs.

Contextual AI is also promising another benefit. In March, the company disclosed that it had developed a set of language models based on RAG 2.0. Contextual AI says that they outperformed several advanced LLMs, including GPT-4, in an internal test that compared the models’ ability to correctly answer questions based on information from external data sources.

Contextual AI says its platform for building RAG 2.0 applications reduces the development process to a few minutes. The platform, which is not yet generally available, will be offered as a managed cloud service and as a standalone application that can run in on-premises environments. Contextual AI has already signed up several early customers including Qualcomm Inc., which uses the startup’s technology to help its engineers troubleshoot technical issues. 

Reuters reported that Contextual AI will use its new $80 million round to bring its platform to market. The company reportedly has not yet set a launch date. Contextual plans to ramp up its customer acquisition efforts in conjunction.

Photo: Contextual AI