Hello, I have an issue with my document interpreter ai (https://beta.pickaxeproject.com/user/control-panel/Document_Interpreter_21DFA?admin=8S4ATITB7FI14MDHXSZD)
I get the issue 'I’m sorry, but as an AI, I’m unable to access or analyze specific files or documents directly. However, you can share the text or the main points of the document with me, and I’ll be able to help answer questions or provide a summary based on that." Even though sometimes it works as intended and other times it doesn’t work.
This is a small eccentricity of how document reading works under the hood.
Essentially, you can’t always say “What’s this document about?” and get an answer.
The system will look at keywords in your response and then search the document for anything that is semantically related and then read it to inform an answer. For example, let’s say your document is an essay on biochemistry and you just ask “what is it about?”, this may not be relevant to the document.
I would recommend going to the advanced settings and changing the “relevance threshold” to really low, like lower than 0.5. (see screenshot)
That was a useful guide but I would like to understand what type of documents are best? I would like a AI to analyze financial data - customer investments that is presented in a csv format.
So the question is what file formats are best? @admin_mike
Text documents are by the far the best. Our document interrogation system works by chunking documents into uniform ~300 word chunks. So chunks that need context sometimes suffer.
Tables are a great example of this because something 500 rows down may lose the context of the column header at the top of the table. That’s why tables uploaded as CSVs perform worse, though we’ll be pushing an update this week that should read them much more accurately.
