The Joy of Discovery - doraithodla/notes GitHub Wiki

Finally I feel that I understand the entire flow of a chatbot, thanks to a RASA article. I will write it down after I get some code written. But here are some of the steps

Actors:

• User
• Bot
• Model
• Vector Database

Building the model

• Read training data
• Clean data (remove punctuation)
• Tokenize sentences
• Word tokenize (preserving word order)
• Vectorize
• Store the text as a set of vectors (and their semantic links)
• Create work collocations

Train

• Update the model with training data (add, delete). Perhaps in the first version, train is the same as building a model

Test

• Test questions and answers
• Correct the weights

Use

• User question(prompt). Initially only take questions (ignore prompt processing)

• Go through the process of clean, tokenize, find matches in the vector database, Find the higest probably answer, display it

• RLHF

• simulate several RLHF sessions

• Update the model

• Version-1: Start with the LM and the chatbot

• Verstion-2: RLHF

• Version-3: Build, train, test the model