====HOW TO USE KAEL==== 1. Train KAEL: Provide some training data first! You can type your own text, copy and paste small stories, or drop in raw, clean text to form its active vocabulary baseline. 2. Choose to use Reasoning Mode or not: Click on the custom Kael avatar on your screen to toggle between Standard Mode and Deep Reasoning Mode. You will hear an audio confirmation and see the background system lists change style! 3. Set your target sentence limits: Adjust the custom slider on your stage window to dynamically choose exactly how many completed sentences you want the engine to build per response. 4. Press enter to chat: Hit your Enter key to bring up the console dialog prompt box, type your query, and hit send! ====TIPS AND TRICKS==== 1. Use Full Text to see your whole chat with KAEL! 2. Wikipedia Optimization: KAEL is fully optimized to read factual Wikipedia articles. For the best logical transitions, remove square citation brackets (like [1] or [2]) from the text stream before starting training cycles. 3. Repetition Breeds Logic: Want KAEL to learn a specific fact deeply? Repeat the exact sentence sequence multiple times in your training prompt block to automatically increase its internal statistical probability maps! 4. Spotting Dead Ends: If you see KAEL rejecting specific word starters while running in Reasoning Mode, it means the algorithm checked ahead and realized that path would lead to broken grammar or a premature cutoff. ====BEHIND THE SCENES ENGINE==== • Tokenized Parser: Instead of processing text as a blind string layout, the custom code strips spaces and maps punctuation (like "." or "?") as independent vocabulary tokens. • Bigram Matrix: Every word connection you train creates numerical scores across parallel mapping lists, acting as a lightweight local memory network. • Look-Ahead Search: Reasoning mode looks ahead past the immediate next word choice to ensure paragraphs maintain structural flow, safely solving the "A cat..." vs "Cat..." grammar issue automatically.
===THE ENGINE BEHIND KAEL=== KAEL operates as a pure, multi-list text prediction matrix. It splits raw text data into token chains using customized string-cutting arrays, analyzes context windows, and uses running weights to determine the mathematical probability of the next chosen word. ===REASONING INTERFACE FEATURE=== If you click the avatar, it will activate Reasoning. This activates the internal focus layer, allowing KAEL to break down user inputs, isolate core subject keywords, and explicitly display its analytical search path before synthesizing the final factual text response. ===INSPIRATION AND CREDITS=== Massive shoutout to Nith04 for pushing the boundaries of Scratch AI engineering and inspiring this project. KAEL was designed as a lightweight, hyper-optimized 3-gram alternative engineered specifically to run lag-free on TurboWarp using 100% vanilla block logic. Visual Assets: Stylized vector logo designed with a modern geometric gradient theme. Typography Layout: Heavily inspired by the clean, architectural fixed-width profiles of Google Sans Code Monospaced. Compiled and built entirely using TurboWarp compilation without external extensions. Feel free to remix, examine the lists, and explore the underlying probability math!
