My team and I are using LlamaIndex's LlamaParse to read in PowerPoint documents. We've noticed inconsistency issues when running the same user_prompt setup for multiple files in one run. We have a small pipeline set up to run through each PowerPoint file in a file path, parse it, do some basic cleaning (grabbing slide numbers, removing trademark tags, etc.) and place information into a pandas dataframe. We're using the premium (45 credit per page) tier. When running these files, we notice that sometimes, images get parsed and described in some files but not all. There are other differences, such as how trademarks notes are described and how bullet points are shown. We understand that use of an LLM will result in new responses every time, but shouldn't the approach be similar? Is there anything we can do to ensure consistency across runs? Continue reading...
