Introduction

I'm going to chat with you today and show you what Llama Pass is. It's something that I stumbled upon recently. I haven't had much of a play with it and perhaps I can inspire you to go ahead and give it a go and have a play.

How many people here are running open claw or have at least heard of that? There's a couple of tentative hands. Wonderful.

This is something, Light Pass, is something that you might be interested in, which we'll have a quick chat about as well. that's something that well I'll just tell you now it's something that you can run locally and you can plug into as a skill and plug into your open claw or other agentic harnessing

software you could do it use it as a skill with clawed or whatever you'd like okay so I'm going to dive right in to start off with creating an agent because it takes between five to ten minutes to run up, and then we'll have a look at LamaPath.

So hopefully the agent will have created by the time that we have a little chat and ask and answer any questions, but I'm not 100 % sure whether that will.

So let's go straight over to here.

Accessing LlamaIndex Cloud and Free Usage

Now, you can access this from cloud .lamaindex .ai. Now, the beauty of this is that you get a very generous credit token fee, which is free plan usage over here, up to 10 ,000 per month.

So if you're running a small business, you can pass your invoices or your documents or do quite a lot at scale.

Or if you want to, do it locally for free on your own machine. And you don't need too much hardware to get that in. You don't need GPUs, et cetera.

Live Demo: Creating an Invoice Extraction Agent

So we're going to do a quick invoice agent. And let's just start off. I'm going to start off with a suggested prompt because it's easy and I don't have to type all that but I'm going to Add a couple of things.

I want the output in CSV you could ask for Jason or you could ask for markdown or whatever you would like and we're just going to press Send and this is going to now go and start creating an agent and I would be able to call that agent

Here we go off it goes. So let's go back to llama pass while that is doing it We'll have a look and see what it's doing first, make sure that it's all looking good. So it's going ahead, it's creating everything.

The impressive thing with this is it's going to do all the plug -ins for us. It's just going to go and vibe code that agent away for us. And I've run it once before. It has been successful.

So here we all go together and see whether it's going to work a second time. We never know.

What LlamaParse Does (and Why It’s Useful)

All right, so LamaPars is something that you can use to parse complex documents and maintain the format of that, and you can get it to output in a number of different file formats. You can even get it to create screenshots if you need an image of a screenshot of a PDF or something like that and extract particular documents, pages from your document.

All right, there we go again.

Okay, so it does multi -passing formats, so you can do lots of different types. It's really exciting, this demo, isn't it? I mean, it's not videos or anything like that, but it is quite useful for the small businesses.

Why this matters for invoice/document processing in the real world

And my background is in AI engineering. I have done a lot of things with generally Azure platforms, and to see this sort of technology become free and available to the wider public is fantastic, because otherwise you would actually be paying for an Azure estate, state and you would have to train a model you would have to go through test that model add documents into it it can be quite frustrating that sort of iteration of work to get out like

invoice processing as you know as exciting as what it is additionally when you're doing that sort of project you have a new supplier come in or they change their format or they change from having the invoice on page one and two of a document to page five and six for example and and I've had that happen. They might not put their logo on. You know, they might do a number of different things that you have no control over what's coming in.

But with this sort of extraction mechanism, you can tweak it a little bit easier, a little bit faster, and you don't have to have a skilled engineer. You can do this yourself very easily, very quickly.

So you get accurate table extraction. You get visual element extraction and custom instructions to do things with, like I added there to create a CSV file for the extract plus you get 10 ,000 free

credits I'm not on any Commission by the way I just stumbled across this and I

Why Use LlamaParse Instead of Chatting with PDFs in GPT/Claude?

wanted to share it with you okay and why would you use llama pass over something like GPT for our because you pro or even five five point four whatever you've got Claude why would you use it well it's quite it's free for up to ten thousand you know tokens we will be using a few tokens in this demonstration but it I mean the price point

you just can't beat it really but you've got a number of reasons so well if you've got one -off documents sure put it through Claude because it will understand your PDFs and you can ask the questions or even notebook LM for example which is a great tool if you're doing lots of PDFs then then you would probably want to have a repeatable pipeline

so you can, you know, share it with people and, you know, up the scale if you need to. You can do, you know, very good chunking of the documents

and what that means is, like, what size of the content will be chunked out. And extracting tables and layout, that's wonderful. It's testable, it's automatic...

Oh, I can't say that right now. But you can automake it, OK? scalable as well okay so let's go back and see what's happening here and see

Audience Q&A: Long Documents, Translation, Handwriting, and Images

what we're up to I guess the first use cases all this policy documents contracts are extremely huge and sometimes very ambiguous as well so if the length of the document kind of increases for maybe hundred pages or 200 pages how accurate the parsing will be so that's one question and second is

like translation from any language to English number three is whether it can

pass handwritten notes and maybe another question is so let's say the document document has images, but in images, the tables are represented in the format of images in some of the documents. So that kind of parsing, is that possible as well?

Because these are all the issues that I face on a daily basis in production. So just wanted to... Sounds like fun. Yeah.

So a lot of times you just got to play around with these things. So the questions are multilingual, the documents are quite technical. you've got to make sure that they're extracted exact verbatim I would imagine especially if they're policies you've got 200 to up to 200 pages that's one question the next one is the

handwritten like how like doing handwritten notes they're very good with handwritten notes nowadays I mean it can even understand my handwriting and I can't once I make notes I just can't that's how it is but you can do a number of things with that you could set up a pipeline to extract the text as as it is with the multilingual.

You might have to chunk those documents, you would have to just give it a go and see whether you needed to do that and see where the boundaries were.

I've had to work with a number of different scenarios and different sizes of documents and you get to a point where the network will, the latency there will just not be as good

as what you want it to be.

So the handwritten, I'm surprised how well these things work with handwritten notes now, they're quite amazing. but I would just try it set up a free account and have a look and see what you can do with it but

Data sensitivity, privacy, and running locally

you have to be mindful that when you're using some of these tools you don't know where your data is going so if that data is sensitive then you probably want to have something perhaps running

on your own server and you could do something with maybe as your this I'm playing around at the moment with another demo isn't about this but with document content understanding and I'm really really impressed. You may want to have a look at that.

It can extract exactly or it can actually generate classifications over a number of those fields if you wanted that sort of production. It depends what you're after in that.

But if you're just trying to scrape the documents, you may have to chunk them first through a pre -processor and set up a pipeline that is 20 pages at a time or something like that and it do you have much of a like what's

your turnaround can you do this in batch as well yeah uh yeah i can do it in batch uh i guess as you said trail and trial and run or trail error might be the best possible way that we can test it it out, right?

So when you said batches, you mean like different documents in the same time? No, what I mean by that, is it like a HTTP REST request or something that you have to get the results back immediately for the user, or is it something that you can do overnight and do it in batches in that way?

Overnight is fine. Yeah, I don't want it to be live or sync, but I'm worried about the accuracies. That's something that I want to check.

Okay, well, we can have a chat after, perhaps, if you want to go into more details. Yeah.

What’s under the hood and vendor lock-in concerns

Having 10 ,000 tokens is great, but what is the tech that is actually powering this thing? And is there an open source version, or do they write about it? Once I start using it, if I like it, if it becomes like a lock -in, then that's a problem. If it becomes like a lock -in, the technology?

Yes. Is that what you're asking? Yeah.

Well, you could use it quite modularly. I don't see, you know, because these things are getting faster and faster to spin up an agent, you just saw how quick that was. We haven't tested it out because I haven't been able to get to that point yet, but we will try.

but because you know you you would actually just make sure that that is a modular section of your pipeline so if it's not a lock into a particular model or anything like that to be honest I don't know what it's running under the hood I haven't really had that much time to really get in there

and have a look I just wanted to share it with you and it was a nice you know let's have a chat about that after as well if that's okay yeah thanks so much for having a look and go and play Play with it.

Conclusion and Next Steps

As I said, it's Llama Pass, the URL, hopefully you got that from the home page, here we go. So llamaindex .ai or cloud .llamaindex .ai and off you go and play, enjoy.

And it's really easy to use. You just saw how easy that was there, there we go.

Okay. Perfect. Thank you very much.