Generative AI for Enterprises

At the end of 2022, OpenAI launched the ChatGPT chatbot, pushing the capabilities of natural language processing to unprecedented levels. Such text-generating technology, joined by other AI technologies for generating speech, images, or animations, has paved the way to a new era of generative AI. More than ever, AI has proliferated into our work and daily life. For an enterprise, AI adoption is no longer a unique competitive advantage, but a necessity for the enterprise’s survival and growth.

The underlying algorithms of ChatGPT are GPT-3.5 and GPT-4, collectively known as large language models. In just one year in 2023, several technology companies successively introduced their own large language models – notably Meta’s Llama 2, the startup Mistral’s Mixtral 8×7, and Google’s Gemini. Among these, Llama 2 and Mixtral 8×7 are open-source models, allowing AI developers to download, fine-tune, and integrate them into various software applications. Open-source models can be deployed in the cloud or on premises, meeting the needs of enterprises for data security, customization flexibility, and computational cost management.

Each aforementioned large language model has undergone extensive pre-training, leveraging massive computational resources and training data. In applications, the user inputs a text prompt, and the model generates new content based on the prompt, via word-by-word prediction untill satisfactory content is produced. However, since generative models only predict the content without judging its correctness, generated content, even if fluent and coherent, sometimes deviates from facts. This phenomenon is known as artificial hallucinations and is a common issue with generative large language models. Furthermore, since the pre-training data of the model mainly consists of publicly available information, the original large language model is not suitable for generating specific content for individual enterprises.

To effectively utilize large language models for relevent content generation, enterprises can employ two methods: customized model fine-tuning or retrieval-augmented generation (RAG). Model fine-tuning requires substantial computational resources, while RAG can use an existing pre-trained large language model, making it easier to implement. Therefore, in enterprise applications of large language models such as question answering and information search, RAG is considered a basic feature.

The figure above depicts the concept of RAG architecture. Before inputting a user prompt into the large language model, it passes through a semantic search system that retrieves relevant information fragments from user-provided documents or the enterprise’s knowledge base. The search results are then combined with the user prompt and inputted into the large language model to generate a response, thereby improving the relevance and accuracy of generated content.

Executing RAG involves many technical details, and there is no unified approach, leading to various RAG solutions in the market with different levels of performance. D8AI provides an RAG solution with the following features:

• Users can select among different large language models such as Mixtral 8×7, Llama 2, or ChatGPT.
• It utilizes a semantic search system developed by D8AI based on deep learning (utilizing BERT).
• It balances content accuracy, computational efficiency, and usage flexibility.
• It is customizable and expandable and can be integrated with D8AI’s other AI-based services.
• It can be deployed in the cloud or on-premises, along with hardware support.

In addition, D8AI also assists enterprises in executing customized model fine-tuning.

Similar to previous advancements in mechanical power, electricity, computers, and the Internet, AI is rapidly propelling our technological civilization forward, and generative AI is an enabler for the current revolutionary progress. We can anticipate that enterprises, large and small alike, will experience the huge impacts, either positive or negative, brought by this wave of AI.