Improve KB accuracy
I'd be curious to hear your opinion about an issue I am dealing with rn....I have been developing a chatbot for a client of mine and during testing from his side..I have discovered lots of potential ways, how he asks questions to the bot, which dont get corectly being found in the knowledge base...
What I have been thinking about, is to somehow restructure & expand the KB with FAQ style files, which would contain these variations of questions and the desired answers (I think it could be worth a try)
However, I am not sure how to use the existing data in KB and which process to use to extrapolate it / restructute it into groups of FAQ sets questions..
I know I can easily use ChatGPT to do that, but on the openai website that gets me only so far with limited usage right ..
I'd be curious if you have been dealing with similiar issues and what process did you went through to make the bot's precision higher.
Thank you very much in advance for any help on this 🙏🙏
2 Replies
+ 2) question
Also I have noticed, the approach lots of people are doing in tutorials on how to do RAG, is sending the the User Question into KB and retrieving the relevant chunks ofc...but sometimes what happene, is that the user asks a question whereas the returned chunks are not so relevant, hence not a good answer is returned...upon which he then asks a follow up question (as we humans are used to) without mentioning his previous one ofc, which then gets sent into KB again - the user then ofc anticipates a better answer, as he asked clarifying question, but the result from KB is actually worse, as its each time only the last utterance, which gets sent into the KB to return the chunks...
I havent seen yet anywhere anyone how to approach this - only thing that comes in mind is to have smth like sliding window of user inputs (smth like last 3 messages), which would get sent each iteration to the KB, in order to return most accurate chunks (I hope I am making sense...)
genetic-orange•16mo ago
Voiceflow keeps the last messages on a memory variable. Maybe you could try including it on the KB query? Or when alongside with the chunks to formulate the answer (if you do it this way)