I’m trying to train the custom GPT based on a massive Telegram group conversation history (converted to TXT file).
My problem is that even though I’m asking the GPT a question which was previously asked in the group I get a general answer rather than specific recommendation (which clearly appears in its knowledge) despite the specific instructions I’ve defined (to prioritize the file for answering Qs and within it to prioritize specific users’ responses) . Does anyone have any suggestions regarding how to solve it?
Just a thought: sometimes it can help in instructions to provide more details about the file that you uploaded, i.e. what content does it contain, how is it structured, how should the model use its contents etc… Specificity can help in such cases although it is not a silver bullet solution. Also, even though it seems you have indicated a priority in your prompt, sometimes you need to be even clearer, i.e. by instructing the GPT to always look for the answers in the file first and only to use other sources if the answer can’t be extracted from there.
Good luck!
1 Like
Hi. I’d like to try and do something similar. Any chance you can share more on the methods used with the Telegram API to make this happen?
You can generate embeddings for these conversations and then create a RAG system which can answer any question based on these conversations which would be ideal for your use-case. There are few 3rd party solutions as well which can help with this