Generative artificial intelligence startup Cohere Inc. is looking to take on rivals such as OpenAI LP more directly with the launch of a new application programming interface that allows developers to integrate its Command large language model into their own applications.

In a blog post, Cohere said the Chat API can be used to build applications such as a knowledge assistant or customer support chatbot in a simpler and more reliable way.

Alongside the new API, Cohere announced that anyone can now test out the capabilities of its new chatbot, Coral, which was built using the Command LLM. Users will need to sign in to the Coral Showcase with either their Google or Cohere credentials before they can access the chatbot. Coral was launched in July, but only for Cohere’s existing enterprise customers.

Cohere said one of the main advantages of its new Chat API is its Retrieval-Augmented Generation or RAG capability. This gives developers a way to control which data sources the chatbot uses when responding to specific prompts. It means they have a simple way to restrict their chat applications to use only the data they supply to it.

According to Cohere, the RAG system helps to improve the accuracy and relevance of generative AI responses by adding data that was not used in its original training. Developers have two choices: They can upload their own datasets, for instance a company’s proprietary data or knowledge base, or they can allow it to search the internet for information.

“For example, a developer building a market research assistant can equip their chatbot with a web search to access the latest news about trends and competitors in their space,” Cohere explained. “We trained Command specifically to perform well on RAG tasks. This means you can expect high levels of performance from Cohere’s model.”

Despite Cohere’s claims, the newly released Coral chatbot did not always provide the most up-to-date responses when SiliconANGLE tested it with a few initial queries, asking about current news events. The response times were also somewhat slower than those of ChatGPT. However, Coral’s actual responses did appear to be accurate, with sources cited to back up its claims.

Cohere said developers can also take advantage of some modular components it has built to aid chatbots built using the Chat API. For instance, the query-generation component instructs the chatbot to only return search results based on the user’s query or prompt, while the document component allows developers to specify which documents it should reference when answering questions. Cohere said it will expand this ecosystem to additional components in future.

Cohere has a lot of pedigree and is viewed as one of the main rivals to OpenAI. Besides the Command LLM, it has also developed more specialized models for content generation and text summarization. Cohere also allows developers to fine-tune these models by training them on their proprietary datasets. In this way, customers can train Cohere’s models to perform more specific tasks, speak in a certain way and so on.

Cohere is led by co-founder and Chief Executive Aidan Gomez, who previously worked as a researcher at Google LLC. He’s best known as one of the co-authors of a groundbreaking academic paper about “Transformers,” which are a special kind of neural network optimized to process text.

They have enjoyed widespread adoption, and sit at the heart of Cohere’s AI models, and can understand and generate prose with much higher accuracy than earlier neural networks. They are also superior in tasks such as writing software code. OpenAI’s GPT-4, which powers ChatGPT, is also a transformer model.

Gomez dropped by SiliconANGLE’s theCUBE in March, when he said we have only seen the first hints of what Transformers can do, and promised that they will become much more capable in future. The full interview can be seen here:

Image: Cohere