With extensive use of AI, especially generative AI, such as ChatGPT, it has become second-nature for most people to get assistance and ideas from such sources. In fact, even larger companies utilize open-source platforms like ChatGPT or Claude. However, since such platforms are trained based on information from all over the internet, one cannot say they “specialize” in a particular field. Thus, to enhance the quality of content that is generated, building a custom LLM on top of an open source model might be the required approach to arrive at a specialist.
Why generative AI is attractive is because it churns out required content fast, has a high level of flexibility and digestibility. So, by building a custom model on top of these open source models, and training that model on the client's catalog, we can achieve true product intelligence. For example, an abbreviation such as 3M can easily be interpreted as 3 meters by an open source model, but our custom LLM will understand that it is the brand 3M.
With attention-to-detail such as this, companies might see an increase in footfall, customer satisfaction, and overall profit. In fact, a custom LLM might even lead them to have a more coherent brand voice, which would only add to the business’s success.
A potential problem that your company could face is that suppliers speak a different language than retailers. A supplier might send you 5,000 items all categorized simply as "Men's Athletic Footwear."
This, however, might not align with your internal website taxonomy, which might be more specific. For example, you might have separate landing pages for "Road Running," "Trail Running," "Cross-Training," and "Weightlifting."
In such cases, even a seemingly small error, such as classifying a Road-Running shoe in the Weightlifting category, might cause you major losses. Customers who will buy these might injure themselves, and then file complaints, and return the products, all because the automated model was not advanced enough to process the nuanced differences between the two shoe types.
This, however, can be worked upon.
Continuing with the example above, while a Running shoe and a Weightlifting shoe might look alike to a keyword script, they would be two completely different products for an LLM.
A fine-tuned LLM will not only look for the word "Run" or "Weightlift." It will analyze the semantic context of the technical specifications. These might include attributes such as material, type of sole (cushioned, non-cushioned, arched, flat, etc), the type of surface it is best suited for (asphalt, mats, indoor courts, etc), and more. Therefore, it decodes the nuance to each of these products through a thorough analysis of their descriptions and features.
Therefore, the future of e-commerce data operations cannot simply be renting a giant model. It's building a smaller, sharper LLM model on top of open-source technology that is trained specifically on your catalog, to ensure tailored and accurate results.