I sure will!
Just to share a few more details - the bot takes in a message and runs cosine similarity; if no response has a similarity greater than a threshold (to be managed as a hyperparameter), the bot takes the previous message as well, and concatenates those last two user messages before the vectorization. If this concatenated message does not yield a higher similarity response as well, the bot asks for more information from the user. (I initially wanted to concatenate even more messages, but this slows down the bot too much)
If the match does meet the threshold, the bot returns the highest-likelihood match (that does meet the threshold), and then asks if this was a satisfactory response. If the user says ‘no’, the bot gives the next-highest-likelihood match (again, if it does meet the threshold), and cycles through all responses that meet the threshold criterion.
The mechanics of it are working alright so far, but I’ve yet to scale up to proper documents, which would probably happen in a few weeks/months