Part 2 Pattern: AI doesn’t understand

 

The chatbot doesn’t understand me is the most common pain point users have when interacting with conversational AI. For traditional chatbots, this is caused by a poorly trained classifier; for other chatbots, it could be from having a bad search mechanism or simply not having access to the right information.

No matter what kind of chatbot you are building, the core capability you need is to understand what the user is asking for and respond appropriately. Chatbots that understand requests give useful responses; those that don’t understand say something like “I’m sorry, I didn’t understand—would you mind rephrasing?”

There are multiple methods for addressing poor understanding, and all of them hinge on identifying what your users want to accomplish and how they phrase their requests. Chapter 4 focuses on how to extract this data from logs and other sources. Chapter 5 shows how to improve chatbot understanding through training classifiers, and chapter 6 adds generative AI to the runtime mix with retrieval augmented generation (RAG). Chapter 7 uses generative AI at train and test time by creating new training and testing data with LLMs.