Gupshup Inc., a conversational engagement platform for marketing and support automation, today announced the launch of a new family of domain-specific generative artificial intelligence large language models customized for functions such as customer support, marketing, commerce, human resources, information technology and enterprise business uses.

The new domain-specific models, called ACE LLM, enable enterprise users to build applications capable of using internal business knowledge to produce natural language conversational experiences fused with industry specific knowhow. These models provide more precise humanlike responses and have a better speed and scale for answering industry questions while doing knowledge work than other models on the market because they have been tuned for those specific industries.

The ACE LLM family has been built on already existing foundational model designs such as Meta Platform Inc.’s Llama 2, OpenAI LP’s GPT-3.5 Turbo and MosaicML’s MPT, but it has been adapted to match the needs of specific functions for particular industry uses. For example, it can be tuned for banking, retail, utilities and more to match the particular jargon, needs and functional requirements.

“Our conversations with brands have led us to understand that there’s an immense interest in adopting generative AI solutions,” Beerud Sheth, co-founder and chief executive of Gupshup told SiliconANGLE. “However, issues such as cost, security, data residency etc. have acted as deterrents holding them back from exploring such solutions. ACE LLM addresses all these issues, enabling enterprises to capitalize on the transformative power of generative AI.”

Foundational LLMs work well enough in general use because they’re trained in a wide variety of text from the internet and thus generate a humanlike responses on a range of topics. However, industry-specific needs fall by the wayside. Industry customers and employees need very particularized knowledge about products or services that requires expertise on the subtleties, vocabulary and jargon that fits their specific market.

“Business conversations need to meet a higher standard of accuracy, context and relevance,” Sheth explained. By putting the power of domain-specific models in the hands of enterprise business users, Gupshup’s ACE LLM is essentially providing them with an expert assistant with deep knowledge of their industry. “They give enterprises a huge head start in building AI-powered conversational experiences. A non-AI world analogy that explains this specialization would be one of a neurologist who is vastly more adept at handling issues of the brain than a general physician.”

A generic LLM would struggle to tell the difference between different types of financial statements because it was trained on a vast variety of irrelevant data from the internet in order to produce human-like conversation. However, a domain-specific LLM for banking, would quickly and accurately produce an answer based on its fine-tuning and provide a far more insightful conversation based on industry knowledge and internal company data.

ACE can be efficiently fine-tuned on business knowledge, giving this family of models a leg up on generic foundational models with its better grasp on internal concepts, terminology and context. It has a better system for specific responses and can provide precise responses and efficient conversations for customers and employees in need of accurate replies based on the knowledge they’re seeking.

The company built in guardrails into the model to corral it, making it focus on proper responses, and when combined with the company’s knowledgebase it can allow users to adjust it for tone, accuracy, and data source controls. Users also get access to a dashboard for auditing, a teach-mode for non-generative responses, automated testing and analytics for monitoring and adjusting the model.

This helps increase trust in the responses of the model. When an industry-specific LLM generates accurate and relevant content it means that it is a more reliable resource. A generic model is more likely to produce less specific content and hedge its answers when it is unable to determine a proper answer.

ACE LLM is available in a lightweight 7 billion parameter size, which makes it easy to fine-tune and implement and goes all the way up to a 70 billion parameter size for more precise applications. The LLM can generate text in over 100 different languages including Spanish, Portuguese, French, German, Bahasa, Arabic, Mandarin, Hindi and English. The ACE LLM models are deployable in the Gupshup public cloud with options for geo-specific data residency or for enterprise-private cloud featuring high scalability.

As for the future of LLMs specialized for specific industries, Steth explained they are on the rise and that it’s an important step forward for conversational AI. “The ascent of vertically trained LLMs represents a significant influence on the evolution of language models,” Sheth said. “These specialized models offer vast potential in domains where precision, context and specialized knowledge hold utmost importance.”

Image: Shafay/Adobe Stock