Faster Training, Better Intents | RAG Intent Recognition: Explained

Key Takeaways

nlu is dead no no gimmicks here nlu is dead Long Live nlu [Music] bye hey it's Pete today we're going to talk about how nlu the backbone of conversational AI since well forever is getting a major upgrade at voice flow an upgrade so significant that it's basically replacing the old system entirely it's like getting your 998 Nokia and then replacing with an iPhone same purpose but wildly different capabilities nlu has been the unsung hero of conversational AI for years every time you chat with a bot that actually understands what you're asking for that's been nlu working behind the scenes and for years we've been meticulously crafting these systems training them utterance by utterance intent by intent like digital bonsai trees printing here nurturing there all while praying that users didn't say something we didn't anticipate and it worked well sort of when users played our rules and said the things that we expected it was magical and when they didn't well let's just say I'm sorry I didn't understand that became the unofficial slogan of the entire industry the problem was never that nlu was bad it's just it wasn't flexible people don't speak in perfectly structured queries they ramble they use slang they combined three questions into one run-on sentence that would make their high school English teacher weak and training traditional nlu to handle all these variations it's like trying to anticipate every possible way someone might ask for a coffee I'm dying for some caffeine I'd like to get a coffee can I get a coffee coffee me got KNE of that Bean Juice etc etc it just meant endless hours of adding training phrases managing intents and then retraining your system and even one weird customer query could send the whole thing sparring into an abyss of sorry I don't understand into rag based intent recognition with embeddings think of it as nlu that actually went to college and got a degree in linguistics instead of just reading the cliff notes it's a real rag to Rich's story from struggling with basic understanding to suddenly grasping the nuances of human conversation here's how it works instead of looking for exact matches or specific patterns this system translates language into mathematical vectors you can kind of think of them as like I guess coordinates in meaning space and if two phrases mean roughly the same thing they'll be close together in this meaning space so what does the weather look like and do I need an umbrella today land near each other even though they share almost no words it's like upgrading from a dictionary to actual comprehension the system doesn't just match words it grasps Concepts okay now first off I want to talk about how this new system is fast and we're talking really fast like training in second what used to take a few minutes an agent with 37 intents and 305 utterances trains in 1 second that isn't a typo it's so fast that training now happens automatically when you test your agent no more waiting around while the nlu crunches numbers you know I kind of feel like this is like going from dialup to fiber but speed isn't everything this system understands Nuance in a way that traditional nlu could only dream of that is if it could dream anyway so let's take this word salad I got a blue sweater last week but it's too big and I was wondering if I can swap it because I really like the color but not the fit and do you have it in a medium ah they want to exchange an item for a different size it cuts through the noise to find the signal no more training for every possible way that someone might phrase a request for you voice blow users this means fewer intense less training and better results it's like having a bot that actually listens in instead of just waiting for its turn to speak your existing intents and utterances will work but you might want to revisit them with this new system you don't need 50 variations of the same phrase anymore the system is smart enough to figure it out with just a few examples and for your users they can finally talk like humans instead of having to learn your bot's secret coded language now we're not pulling the rug from under you both systems will run side by side for a while so you can test out and transition at your own pace just look for the intent classification toggle in your settings but let's be real once you try the new system going back to traditional nlu will feel like trading your smartphone for a carrier pigeon so is nlu dead at vo flow yeah I'd say it is it's pretty much on its last legs and We Salute You traditional NL for you have for the most part served us well but now it's time to make way for something better an approach that lets us Focus Less on training Bots to understand humans and more on designing experiences that actually help them and that my friends is that an obituary for The Voice flow nlu July 2022 to 2025 let me know your thoughts in the comments and remember as always stay curious [Music]