I am making a “personal assistant” chatbot to send to people who do not know me which will answer questions about me. I am doing this by including a lot of information about myself in the prompt. However, this causes obvious issues as the AI has to tokenize this every single time a question is asked to the chatbot, which means I am using a ton of tokens. Is there a way to just tokenize it once and have the AI remember it, or some way to teach it about myself without having to use tokens?
I apologize in advance if this is an elementary question.
You would be better served (more optimal way to spend time and money) to just build your own full-text based search engine in front of the API. All you need is a full-text search capable DB and a simple way to enter your questions and answers into the DB and create the full-text search indexes and you are basically done.
Databases do great full text search and retrieval, so you don’t need to waste money (and time) fine-tuning for an AI GPT application which can easily be done with a database of questions and answers, etc.
If you really want to experiment with OpenAI, then you should consider doing the same thing described above, except instead of using full-text search, you can use OpenAI embeddings and store the embeddings (vectors) in the DB and then use a little bit of linear algebra to rank the replies based on the search string (which you also convert to an embedding vector).
I have a demo embedded vector search application which does this using OpenAI API generated vectors, BTW:
I am trying to make the exact same thing. Seems absolutely unnecessary and overkill to have to retokenize every single new thread instead of letting us ctrl c ctrl v conversations for this very purpose. Also, the guy above didnt seem to quite understand what you were looking for. Quite frustrated. If you want to work together on this, let me know!
