Hi Everyone! Finally decided on a real pipeline for Atlas’ intent engine. I wrote off fasttext before for being unintelligent. Fasttext doesn’t know real language like that audio and sound are related to each other. MiniLM or SnipsNLU might, but they would be so heavy. The end goal of Atlas is to be able to run on not just iOS, Android, and Linux, but also very low end linux devices like a first gen Raspberry Pi Zero. The intent engine stage should not be any source of bottlenecks or slowness. Fasttext is doing quite well. It kept failing, but with more training data, parameters, and random flags in the training step, it’s not half bad! The only major problem that I’ll run into is with arguments. For example, when you say “What’s the time in Tokyo?”, it can figure out that you’re asking what the time is pretty fast. The problem is making sure that it passes along “Tokyo” as an argument (or entity as some call it), or else it will just completely fail. This step is hard, but keywords should be easier to sneak in if I am desparate. Keyword based intent is very very unreliable though, so I’ll have to see and do some more research on fasttext’s potential. For now, it’s quite nice. If you want to use fasttext yourself, you just clone a git repo, compile into a binary, feed it one file with your examples and intents, and watch it fail and fail until it doesn’t. Massive respect to FastText devs (even though it’s Meta eww, it’s open source), and happy coding everyone! Ps– it’s unlikely that an Atlas version will be completed during the duration of stardance, but I’ll try. Thanks for reading!