Welcome

Good afternoon, everybody. Thanks for being here.

Do you hear me well? If not, I can use the microphone.

Great.

I'm one of the co-founders with Jorge of SAPin.

85% of AI projects fail due to poor data quality.

We also view a lot of people using ChatGPT inappropriately and wondering if they are doing something wrong. That generates a sensation that something's missing here. A lot of hype around the AI, but I am doing something wrong.

Why does ChargeGPT doesn't read my PDF? What's happening here?

Well, the problem is that that person probably isn't aware of the limitations of ChartGPT when it comes to process both PDFs, Excel and so on.

And that's what we are going to discuss today.

Let's start with a demo. I'm gonna show you first of all You all know this interface, ChatGPT, of course.

I am going to show you a comparison, a brief, brief comparison between how ChatGPT failed to read information from my PDF and why, and on the other hand, how can we bypass that limitation and give ChatGPT the information that we want him to process and do it properly. So, first of all, here we have this PDF. This PDF is the one you are seeing now.

The content doesn't matter. What it matters is that for any given reason, I want to read this information. I want ChatGPT to read this slide, the next one, the other one, and so on till the fifth.

Okay, so now I enter 2jgbt, I upload my PDF, and I give it a little and simple prompt. Okay, I have an answer, but is this answer correct? It's not, right?

In the document, we are seeing stat-free defined implementation requirements, but on ChatGPT, we don't see this. What is happening here? Anyone wants to give us a reason? Because... Okay, okay. Great, that's it.

Just to show you, I've done the same process using exactly the same or a let's say a approximate prompt, but in this case, I have uploaded not the PDF, but the screenshot of the slide, and in this case, ChargeGPT has read correctly the information, so it has given me the correct information.

We'll see much detail later, but the reason is just right there. ChatGPT it's not able currently to read images graphics from PDFs It's just only able to read text from them. So that's why he wasn't reading properly my information Then let's change.

Let's move from PDF to Excel Excel here we have an example of in which I've uploaded this Excel, as you can see. It's in a spreadsheet with four sheets.

One of them is this one, which is a little bit bigger than the other. It's more complex and it has some formatting, as you can see. Here we have some merge cells, which is really common, right?

And I have uploaded to ChargeBT. As you can see, TGPT is able to open it up and I can move from sheet to sheet by this dropdown just here.

And, well, when I ask for describe the guidelines two to four, which are there, He does it correctly. It's not failing, right? It's doing its job, and I'm satisfied.

But what happens when we try to use O3, which, by the way, is one of the reasoning models that ChargePT offers? O4, it's not a reasoning model, and O1, O3 are.

When we do it, you see that ChatGPT opens up the spreadsheet, but it's not able to answer my prompt. Why? Because all three models, the reasoning models, aren't They don't support both Excel. They don't support currently Excel files. So if you need to do anything regarding Excel files on CGPT, don't use all three models because what you will receive is a not good response.

1Let's go back to the presentation, and now that we have seen some visual representations, let's go a little bit in detail around current limitations around PDFs and excels on chgbt, as they are the most common files on a daily basis that we use. The first is Excel. As I just showed you, all three models, all one, they don't support currently Excel. They support images, docs, PDFs, but don't Excel.

When you use, it's important to know, this is an interesting fact, which is, Let me look for it. It's here. I change it. Let's look for it.

Okay, I don't have it, but when ChartGPT reads an Excel file, it's using a Python library to do it. Practically, that means that ChatGPT doesn't treat Excel as an interactive interface. It extracts the raw data.

What about PDFs? Very briefly, CharGPT, as we have seen, just processes text from the PDFs. It's not able to read images, it's not able to read graphics.

When uploading a large document, and we are talking about around 200 pages, the model might only read one part of the document. because it has a context window which is his short term memory and if you fill that memory the chat will respond to you but just using a part, a fraction of the whole document. What's the tip here?

If you want to process a big, big, large document, you can chunk it. You can separate that document into three or four parts so ChatGPT is able to process it completely.

As we related, it doesn't visually interpret the PDF. It's not like he's able to read this whole PDF, for example, and he's able to read the images. He's only able to read the text from the documents. So that's an important part.

without clear delimiters, the model could mix up columns. What does it mean? Here, as humans, for us, this image, it's really clear that we will start to read it from the left, and maybe not from the bottom, but from the top, and it's like really intuitive to us to read this page.

But what about LGBT? Probably, if the document is not clear, he will start from this side and then he will add up a column from this side.

So we could have a mixed up of content from a page that could confuse the model into combining information from places that it's not where we want to combine the information from. So just take that into account.

Finally, regarding PDF files, the size limit is around half gigabyte, 500 megabytes. It's a lot of size, but just take it into account if you work with huge and large PDFs.

Best practices to sum up.

Regarding PDFs, prefer text-based PDFs. That's a mantra that I have when working with generative AI, is that when I work with a document, I want it to be as close to plain text as possible. If I can have a Word document with just text, I prefer it obviously in front of any other document.

What's our job as people who want to use AI? Appropriately, we have to transform documents.

If we have a PDF with a lot of images and graphics, I recommend to create a Google Docs, a Word, and to keep moving or keep pasting the text of that PDF into a Word, into a plain text document, and then we upload it to ggbt. We simplify the layout, we chunk large documents, and always verify the extracted content, as we've seen in the examples, because ChatGPT could hallucinate and create content that was never into the PDF, but if we don't check it, if we don't double check it, we could be entering into a misleading information.

And regarding Excel, I recommend to use one record per row. It's very common to have an Excel From row one to 50, some information. And then from row 60 to 80, some other information.

Regarding Excel, the key here is to format and keep the spreadsheet as simple as possible. So we have just one line of information per column and the same per row and we upload a sheet per sheet, one sheet at a time to ChatGPT.

Thanks and let's move to the questions.