VectifyAI/PageIndex ? reverse-engineered prompt

I'd like to build an intelligent document analysis tool in Python. I'm finding that current AI tools for asking questions about documents aren't smart enough for complex, long PDFs like financial reports or technical manuals. They just seem to match keywords and miss the real context.

I want a system that works more like a human expert. It should first analyze a document to create a smart, structured index—almost like a detailed table of contents that an LLM can understand. Then, when I ask a complex question, the AI should use this index to reason its way to the correct answer by navigating the document's logic. This will provide answers based on true relevance, not just word similarity.

The system should show its reasoning and cite the exact pages where it found the information. Let's build this without relying on a vector database or arbitrary document chunking. If possible, it would be cool if it could work directly from the PDF's page images, too. You can use OpenAI for the reasoning part.