CustomGPT comes with a bunch of site-wide setting variables that allow you to customise the site by just changing one value instead of many in workflows. Here’s a guide for what each does.
Default 250. The target chunk size for memory. Larger chunks include more information, but use more tokens. Smaller chunks use less tokens but may not include all the necessary information to give a good answer.
Default 1500. When splitting into chunks, if this number of characters is reached without the chunk splitting, it will be split, even in the middle of a word.
Default 20. Measured in minutes. The memory item will be checked this number of minutes from when the upload was started to see if there was an error in the upload process. Do not set this too low, because if it checks for errors before the upload is actually complete, it will mark an error even though it’s still working on it.
Default 0.75. The required similarity score to be given to the chatbot for its answer. For example, if the Pinecone API request returns 5 memory chunks but only two have a similarity score equal to or greater than 0.75, then only these two will be given to the chatbot to use in its response. Setting higher decreases hallucination but may incorrectly remove potentially useful memory chunks. Setting lower increases the probability of it being given irrelevant memory chunks that cause it to get sidetracked.
Default 3. The number of memory chunks returned. Higher gives more information, but uses more tokens. Less gives less information but is more token efficient.
Default You are CustomGPT Support. You answer questions about the CustomGPT Bubble template using the information provided.. The prompt given to chatbot in every message.
Default Here is some additional context to use in your response. Some of it might not be relevant to the question, so only answer the user's specific question, nothing more: The additional prompt given to the chatbot when relevant memory is found.