Cohere Inc., a well-funded artificial intelligence startup, today introduced a connector development tool for its flagship Command large language model.

Customers can use the tool to build integrations that enable Command to retrieve data from external applications. That data, in turn, allows the model to answer user questions it otherwise wouldn’t be capable of processing.

Toronto-based Cohere is led by co-founder and Chief Executive Officer Aidan Gomez, one of the eight researchers who invented the Transformer neural network architecture in 2017. The company reportedly achieved a $2.2 billion valuation in July after raising a $270 million Series C round. The investment included the participation of Nvidia Corp., Oracle Corp. and other major tech firms.

Cohere offers a collection of LLMs that enterprises can deploy in the major public clouds platforms as well as on their on-premises hardware. The company’s flagship model, Command, is capable of generating text, searching document repositories for specific snippets of information and sorting business records by category. Cohere says the LLM was trained on a dataset specifically designed to optimize its performance across common enterprise use cases.

By default, Command can draw on the information in its training dataset only when answering user questions. The connector tool that debuted today enables the LLM to access information from practically any external system with an application programming interface that supports data searches. Moreover, the tool works with the cybersecurity controls that are implemented in external systems to limit users’ access to sensitive records. 

Cohere has released about 100 prebuilt connectors on GitHub to ease the development process for customers. According to the company, those connectors enable Command to retrieve data from popular cloud services such as Asana, Slack and GitHub. There are also prepackaged integrations for other types of applications, including popular open-source databases.

The new connectors build on an existing RAG, or retrieval-augmented generation, feature in Cohere’s Command model. RAG is a machine learning technique that allows AI models to access information from external applications without major code changes or retraining. The technique involves pairing an LLM with a second neural network that turns data from external systems into embeddings, mathematical representations of information the LLM can more easily understand.

Cohere says Command’s RAG feature includes citations in the answers that it outputs. As a result, users can quickly find the documents on which the model drew to generate a response.

“Companies can now ground their answers in virtually any proprietary information that they have, regardless of what third-party app they are using for corporate collaboration,” Cohere engineers Roy Eldar and Beatrix De Wilde wrote in a blog post. 

Several of Cohere’s competitors also offer the ability to extend their AI models’ knowledge base with external information. Notably, OpenAI is currently testing an API that enables customers to fine-tune its flagship GPT-4 model on custom datasets. The company offers an identical capability for five of its earlier language models.

At its DevDay developer event last month, OpenAI launched a program that enables companies to customize its models even more extensively. Participating customers can modify “every step of the model training process” to optimize the company’s LLMs for their requirements. Each customized LLM developed through the program will only be available to the organization that commissioned it. 

Image: Cohere