Let Your Customers Chat to Fill-Up Shopping Cart

While an LLM model sits at the core of the product, we found there are multiple steps Sauja needs to handle:

* Input cleanup: The user is asking in a specific language, cultural context, prior shopping history, and product preferences. Most of all, they hold assumptions about the world. This needs to be conveyed to the model in the right way.
* Multiple scenarios: The model needs to recognize one of X scenarios – listen to the user, suggest ideas, talk in detail about a suggestion, be brief, ask questions, and narrow down the choices. Models such as GPT-4 are robust but need tweaking in this regard.
* Model actions: The model needs to be taught what actions it can perform. The user sees a ChatBot. Your back end sees an agent that searches, filters, plans, thinks, and yes – talks to the user.
* Indexing and search: We need to index your data in a way where it’s searchable by the model.
* Model output refinement: What the model produces is not the ideal search term. Brands, prices, preparation steps, quantities, product promotion & placement, quantities, stateful variations (powder garlic vs garlic), cultural & language specifics – we augment the data to help the search.
* Output guardrails: The output then needs to be cleaned, trimmed down, and structured so the code interface can read it.
* Finally, what kind of UI is needed for the site? Which way do people prefer to interact with such a system? We carry on the research and tailor the UI elements to your site.
* How do customer insights reach you? How did they engage with their cart? Did they regret the search results? How much time did it take? What exactly were they looking for? We strive to make sure you learn what your customers desire.
* Latency: We’re well aware that models take a long time to run. We work to make this imperceptible, or acceptable for the user.
* Updates, infrastructure, security: This is work-in-progress. Let’s chat, and figure out whether cloud, on-premise, or another combination of infrastructure works best.