Modernizing With HCL Informix: Bringing AI to Your Existing Data With VectorBlade

Hi, everyone, and thank you for tuning into this webinar. I go by JC, and I'm Chief Architect for Informix. For those who don't know me, I began working for Informix in 1991, a magical time when dinosaurs roamed the Earth.

I was hired into frontline support, not as a programmer. So when I later moved into development, I'd come by way of all levels of support. Obviously, my job now is to focus on innovation.

But I like to think that after spending a good part of the mid-'90s wearing a beeper on my hip, resurrecting down systems all manner of day and night, I really do care about software quality as much as I care about innovation. A major innovation Informix made back then was extensibility. Of the big four databases, Informix, Oracle, DB2, and Sybase, Microsoft wasn't the player it became in a couple years later.

But of those four, IVS was really the first to allow an outside developer to create their own data types and their own C code or Java code to run inside the server as a first-class citizen. We called each related bundle of user-defined types and user-defined routines a datablade, and lots of developers went to town with this feature. We did, too.

We wrote a handful of our own datablades that are still available and popular 30 years later. When you download IVS in 2026, you get bundled with it a basic text search blade, a geospatial blade, and a time series blade. Now, early this year, we set out to make IVS a great place to natively store and use vectors, which is a complex data type that isn't technically new, but its criticality to the world is fairly new.

And right away, we realized that the best way for us to do that was by leveraging our own extensibility technology. So five months later, here we are. We've done that.

And today we are unveiling the HCL Informix Vector Blade. Now, right out the gate, I'll mention that this blade does not work with an IBM Informix backend. It only runs with HCL Informix, and it is not supported on any server version lower than 15.0.1.

I'll tell you the exact minimum version we recommend a bit later. Meanwhile, we've got a team that is keen to introduce it to you, so we'll get started. We have data strategist, IBM champion, and longtime Informix champion, Jean George-Perrin, who goes by JGP, to talk about what vectors are and why they are so useful in artificial intelligence applications.

Then I'll swoop in and talk about the power that our blade gives you over this new world, and we'll have a demo by Informix field engineer and all-around excellent guy, Scott Norris. Now, while all this is going on, feel free to ask questions, and the team will do our best to field them live. Just use Zoom's Q&A feature and window rather than the chat, if you would, because that's where we'll be looking for them.

So thanks again for joining us, and JGP, please take it away. All right.

Hi, everybody. I've been in the Informix world for a little bit as well. Okay?

So not as long as JC, and probably not in a prominent way as JC, but I started my journey with Informix back in 1997 with a company called 4js Development Tools. And, it started-- Something I like to say about Informix is that before that, I only had college level knowledge of databases, and I really hated them. And karma hits, and it stick to me.

So when I joined Actian a year ago, almost to the date, it's been really an interesting moment to be back in this fantastic Informix family. And I'm really happy to see that this webinar is pretty well attended for 1:00 a.m. in Sydney.

Okay. So why do we do vector? Okay.

So the whole idea is that vectors are what is driving similarity, and this is a best way we've actually find to do that, at least for now, when it comes to comparing different things. Okay? Comparing text or like, a puppy is a dog.

And this is a kind of similarity we need to do to understand better the prompts and the questions and the semantic approach to doing AI. So There is actually two ways to do it. The algorithms themself, as we're going to see, are not new.

And it's always the same thing, like big data allow to do more analysis. Now we've got the compute power to actually do AI like we were always dreaming to do. Okay?

So there's really two ways of doing that. The current way, or the popular way, is to actually do it in a NoSQL database, to install some kind of a vector database you would download, you would install, you would manage separately. It's a stand-up.

You've got to build and maintain pipelines, and we all know that pipelines are brittle by nature. You've got duplicate data. You've got to have to learn a new system and a new security model for those databases.

And of course, part of the life cycle of all of that, you've got to do backups, and of course, you've got to provision the infrastructure. So this is how people are doing it right now. Or you trust what you have.

And if you're using Informix now, it's probably because you already trust it for a long time, and you can just do all these exact same operations as you were using the engine. Okay? That's a benefit of what JC was just saying about this opaque types that are really part of our strategy.

Next slide. So, and we all know, okay, we all know that AI is difficult, that the ROI around AI is not easy to find. We also, when ChatGPT started, the magic of I am asking a question, and I'm getting a good answer.

I'm asking code code to generate some code, and I'm getting good applications. But when it comes to the enterprise data, to the enterprise knowledge, it's actually very difficult to get something, okay? And those figures are kind of scary.

Okay? You see that MIT saying that GenAI delivers no ROI, 95% of them. You see that McKinsey said that deploying agent is very difficult.

You see that Gartner is saying that 40% of the program will be abandoned, okay? So there's really a lot of threat around this thing. The challenge is there, do you really want to adding a new database or a new database strategy on top of that?

So, that's really the kind of questions you can ask yourself, okay? And if we look at the iceberg, the iceberg we've all been seeing. So, 20% of the data is structured, and that's an estimate that has been floating from Gartner for a long time.

And the 80% of the unstructured data. Okay? So there was always a myth that we could be able to transform the unstructured data into structured data, but this is not happening, okay?

And the thing is that nobody's doing it. So one way to do it is to actually put all those unstructured elements, all this unstructured text image, et cetera, into a database where you can actually seamlessly kind of ask to dig into, like you would do your normal SQL. Okay?

So, that's one of the value of that as well. And as JC was saying, the BLADE technology that has been at the heart of Informix and the design of the architecture have allowed to support a lot of different models. Okay?

I remember a lot of efforts around time series, and that made Informix a very popular choice in the telco industry. Or even Informix being one of the first engine to support natively XML. So that was when XML was popular.

I still like XML. So all of that really meant that you're not choosing a database based on the current trend of something. Okay?

Like, "Oh, we're doing vector, so let's have a new database for vector." "Oh, we're doing relational. Let's have a new database for relational." Okay? So we're really being able to capitalize on the existing assets, on the existing data without movement, et cetera.

So that's really why now you can have this new eighth data model in this architecture, this Informix architecture. And you see it, okay, when you compare it to different products, okay? Here, I just compared four popular databases.

So you've got ours, our beloved Informix, and then you see the other guys, okay? So it's always We always thrived by doing the technical and the technology things correctly because of the architecture that was there. Okay.

So for example, when you're talking about replication, replication is almost coming out of the box, whatever the type is. And it makes a difference. So yeah, some are open source, some are free, but it's not only about the license cost here it's coming, it's also coming about the cognitive load to understand how those new products are actually working.

And really when you think of it, it's not a new game for Informix. We started, and I remember playing with Excalibur back in the day. It was probably, like a lot of things with Informix, ahead of its time and a little bit before we added the real use cases and the very popular use cases.

So now with this new wave of indexing, and we've managed to actually come to be a very competitive vector architecture for your data needs. Okay. So I think we've achieved that.

It's pretty impressive and I'm kind of looking forward to a little bit sharing with you and understanding a little bit more about the technical details that JC's going to go through and the demo that Scott is going to do just right after. So JC, back to you. Okay.

So as I said earlier, extensibility is something Informix does really well. And to be honest, adding a native vector data type to IDS took about a day. It's simply an opaque UDT of variable size, depending upon how many dimensions your vector has.

And it's stored internally, kind of as you expect. It's an array of four-byte floats preceded by a number of dimensions. We've sized it so that currently the maximum number of dimensions is 1536, which coincides with OpenAI's text embedding three small vector format.

But if there's demand for larger vectors, we can easily upgrade the type, which by the way, is called Lvector_embedding. I'm very curious to see how people use this blade and what they decide is missing. We have some very creative, clever customers, so I have no doubt we'll have some interesting improvements to make over the next six months anyway.

Creating that Lvector_embedding UDT and a handful of supporting functions. You need two casting routines that will convert from an VARCHAR to a vector and back. You need input/output routines, network send and receive.

That part was relatively easy. I don't know if it took a day, but it was probably close to a day. It's indexing vectors with thousands of dimensions and searching on them efficiently that's a little more involved.

So rather than reinvent the wheel, we decided to use an open source library for a lot of this heavy lifting. Google has one. There's one from Microsoft.

But in the end, we went with LanceDB. Now LanceDB does exist as its own standalone database product, but this is not that. They also have a C library that supports a large subset of the standalone product's functionality, and that is what we incorporated into the blade.

This library is already quite mature and offers a variety of index types such as IVF-flat, and an interesting combination of IVF and HNSW. And of course, the three main distance calculation methods, which are Euclidean, Cosine, and Dot Product. So we combined the LanceDB library with an interface layer of our own, consisting of about 30 UDRs into one shared library called Lvector.bld.

Like all blades, this guy gets installed beneath the Informix extend directory. Now, once the blade is installed, as long as you're running HCL Informix version 15.0.1.10 or higher, it will auto-register. You can use it with .9, maybe even before that.

But it won't auto-register as nicely in those earlier versions of 15.0.1. And what that means is that when your application calls one of those 30 UDRs for the first time for a particular database, the engine automatically does a couple things. First, it runs this objects.sql script, which is part of the blade, and the definitions for those UDRs and the Lvector embedding UDT are added to the system catalogs.

Then if you don't already have an Lvec class VP running, one will be spawned automatically. Obviously, we recommend that you add a line to your config file, a VP class line that includes an Lvec class of some number of VPs that you want running. But initially, you probably won't have one and so we'll automatically spawn one for you.

And then the Lvector shared library is loaded and dynamically linked into the Lvec VPs. At that point, your blade is active and ready to go. So what can you do with it?

Well, here's one possible scenario. You can create an Informix table that has a column of type Lvector_embedding, for example. This table will also have a serial or a big serial column that will uniquely identify each vector, and that's pretty important.

And then it has an varchar column called metadata. The column actually has to be called metadata. That can store any text, but usually you'd put categories in there or keywords that you might want to filter on.

Beyond that, the table might have columns that contain whatever object was embedded, like an varchar column for a chunk of text or a blob column, or something along those lines. This is just a simple example. So you can load thousands or millions of vectors into that table, identifying each vector using the serial column.

You can then index and search on those vectors for the nearest neighbors by creating a column or Lance table and flash synchronizing it with your Informix table. This will very quickly copy all the vectors and the IDs and the metadata into the Lance table, which can then be indexed and searched. You can search the Lance table by itself.

Sorry, Nick, we want to go back. We're not at the next slide. Sorry.

Go back a little bit more. One more. Okay.

So next slide, please. Here's where we are. This is where we are.

So you can search on the Lance table by itself, or you can, of course, join that table with the traditional Informix table or any number of Informix tables to retrieve, say, X nearest neighbors to a given vector, plus whatever text or image or other object is associated with those neighbors. You can do hybrid queries. In fact, you can do them at multiple levels.

For example, the Lance table will store the metadata along with the vectors, as I said, and this metadata can itself be indexed and filtered on. You can also filter on one or more columns in your Informix table if you're performing a join. The purpose of this would be to get the 10 nearest neighbors to a vector based on both semantics and some keywords, for example, but then to call that list of neighbors by filtering on whatever you like from the Informix table or tables.

This honing using hybrid queries can significantly reduce AI hallucinations, which we know is an issue, and it's dead simple with this Vector Blade. So where do you get all these vectors? In what part of your application does the embedding happen?

Well, it's up to you. You can bring your own embeddings, and IDS will happily index, search, and store them along with any associated objects, like images, sound, text, whatever, the way it has traditionally done that. But as a convenience, we've also included it in the blade, a UDR called L vector_embed, which will take an varchar as an argument and go out to OpenAI, embed that string, and return a vector with 1,536 dimensions.

The only prereq is that you have an OpenAI API key and provide that key to IDS with a new configuration parameter. Next slide, please. So here's an illustrated example of one specific RAG use case for the blade.

Your organization has a bunch of proprietary documents. It could be technical specs, internal policy docs, anything that's important to you or your organization that you don't want out on the wider web. You feed those documents into an application that chunks the text, meaning a trained model breaks up the document into overlapping sections.

Each chunk is then embedded by a model of your choosing, and you use the Vector Blade to store and index the resulting vectors along with their chunks in IDS. Next slide, please. Now it's time to search on those vectors.

So let's say a user types a natural language prompt into your app, which embeds the prompt using the same model. You pass that resulting vector to IDS and ask for, say, the 10 nearest neighbors. You don't want vectors back.

What you want is vector IDs, which you can then use to find the associated chunks in an Informix table. You then pass those chunks of text to an LLM, which formats the answer, which is displayed for the user, that kind of thing. Next slide, please.

Now, the great thing about this blade, which JGP touched on, is that it's just one fully integrated slice of the IDS pie. All the action I've described here could be happening in one corner of your production database without dragging the performance of any other production tasks. You don't need a separate database.

All your relational data can join in on the fun, and everyone gets along. Next slide, please. So here's what you can do to get started.

Presumably, you have HL Informix 15.0.1.10 or higher already installed and running, and you've downloaded the blade TAR file. It's really just a TAR file that you install into Informix to extend. You configure a few things, such as your OpenAI API key if you're going to create embeddings using the blade.

You then register the blade for your database, and off you go. You can manually register the blade or you can let it auto-register, as I mentioned before. In the Informix Direct Extend Directory, you'll find a sub-directory called samples, and in there we've provided some sample SQL scripts that you can review and execute even.

And at that point, you could start experimenting with the 30 or so UDRs using DB access. And from there, you can begin writing your application. Next slide, please.

So this is a more involved workflow diagram that brings up several additional points. Before you bring your application online, you've probably ingested a bunch of data. That phase is depicted at the bottom there.

You've created vectors for text and images and loaded those embeddings, along with their associated objects, into Informix using the blade. You then index the vectors. Now, when you create an index on a Lance table containing vectors, that doesn't render the table static.

You can insert and delete vectors just fine, and your searches will still make use of the index. So after the ingestion phase, you start your application, and a shopper searches for a product using a natural language description rather than a specific product name. "Show me your least expensive mountain bikes that have disc brakes," for example.

That prompt is embedded by your chosen model. The resulting vector is then used in an ANN query against Informix, which returns the IDs for however many nearest neighbors your application requested. Those IDs are used to join to find the most relevant items in your catalog, and then the list is ordered by cost.

In theory, all that can be done in one hybrid query. The results are then sent back to the user, or you could also use an LLM to format and augment the results before displaying them. That's pretty much what we have, and now Scott is ready to show you a demo that models the same thing.

Take it away, Scott. All right.

Thanks, JC. All right. Let me share my screen here.

Nope. Go ahead to the next-- Actually, Nick, go ahead to the next slide. Sorry about that.

Okay. As JC mentioned, we want to make sure that we are running Informix 15.01.10 or higher. Right now, we have done a lot of testing on Red Hat Linux.

You want to go out to your OpenAI and create a key. People ask if there's a cost associated with that. And you could get a free key.

I think I paid about 10 bucks, and you get a lot of embeddings for that. Make sure you have a lot of space. As we ran through this process, I bumped up against those two gigabyte levels.

So two gigabytes and higher would be preferred. And then your archive. Next slide.

Okay. So as JC mentioned, how you get through this is you get the TAR file from us with the vector blade embeddings. You export all these commands into your shell, and then on your on-config, you will preload all this information.

And then, as we get into it, you will register the engine. So go ahead, and next slide. And here's, again, a couple of sample SQL scripts.

I'll make a note. We're cleaning this up, so please don't go verbatim by these slides. And so we'll send out an updated slide deck afterwards.

Next slide. And a couple just generic how we do some of the embeddings. Again, if you notice any of the UDRs, start with the L vector, underscore, and then whatever that command might be.

All right. And next slide. And next slide.

And now let me go ahead and share that demo. Okay, I am going to share my entire screen here. Everyone able to see that?

Okay. Thank you. Yes.

All right. So first, we mentioned that we do have a number of samples associated with the blade. So I like to be straightforward with everyone to take a look at some of the samples we set.

So basically, setting up the blade is we're going to create a database to put all of our new information into. We're going to execute the prepare statement. This registers the blade for use for a blade.

And then we'll go through several create statements. So let me do this. So we will go ahead and do that.

Everyone cross your fingers that the database gods are nice to me today. You notice we have a couple databases out here, so we'll just select one, and then we're going to go run some of our samples. So first is just the setup.

And so we are running the setup, and we'll go ahead and run that. And basically this is kind of putting all the information into the blade, registering the blade, and getting everything set. Next, I am going to set my key, and again, this is the OpenAI key.

You can do this from your configuration file. What I've noticed is every time I've created a new database, I've always had to kind of reset that key or make sure it's associated. So I just go ahead and associate that key, and I'll go ahead and run that.

And by doing that, that automatically registers the blade for use. So now we're ready to use that. So one of the things that we just completed is we set up all the UDRs into the blade.

So let's take a look at that. Of course. Thank you, demo gods.

There's a missing space after the word from. Yeah, okay. The word from.

Oh, thank you. And where as well. Huh.

All right, so- I think the line wrapping got messed up. Yeah. That's how you see it's live, the debugging SQL from- Ah, yes.

This is what we can do for a living here. All right. So basically, this is just going to show us all the LDRs that are associated with this blade.

As JC mentioned, we have several in there, your normal create tables, build, and sync. So you're no longer sharing your screen, Scott. Really?

All right. We're seeing your lovely mug. Hmm.

I see Scott. I see his screen. Yeah.

All right. He's trying to show us his demo, I think. No, do you see the screen?

You don't see the screen, then? Yeah, I see DB access and- Okay. Yeah.

Oh, really? Yeah. I do not see it.

JC, you're throwing me for a loop. I guess, sorry. For some reason, I'm only seeing you.

All right. No worries. All right.

So let's get out of here. And we're going to just do a basic demo. So basically, what we're going to do here is we're going to create a table.

We are going to execute a create a lance table or a vector table. This name is Demodocs. And just a word of caution, if you happen to run these and you tend to run these over and over again, if you remove the database, it does not remove any of the vector tables, so you have to go ahead and drop those as well, just so that you have a kind of clean slate.

But that's things that I've run into as we went through this. All right. So it's already existing.

And here we go. Let's see. Shoot.

After playing around with this all that, and you... Not cleaning up after myself. All right.

Anyways. So what this actually did for us is created several vector inserts into this, into our vector table. So we have a couple of words on numerics, and we're doing that and committing our work.

So if we select from that table, we should see our closest entry points, which in this case is only three. Now, of course, I'm going to mess that up again since it's on there. Great.

All right. So much for setting this up ahead of time and... Yeah.

Let's see. I do have it down here Sorry about that. That's the other one.

Okay. Hang on a second. Well, like I mentioned, I do...

I did. All these are created, so that's why it's bombing out on me. All right.

Well... One thing that is interesting to note while Scott is going through that is not, in this scenario, Scott is creating his own embedding, so you're not using the OpenAI keys there. Okay, so- Correct.

So in this scenario, that's why you see the vector and you see this limited string there. Right. And you don't have to use the OpenAI at all to use the blade.

It's just- Yes ... provides convenience. Mm-hmm.

None of these are running for me. So what are we doing? I apologize for that.

Got into a wonky state. Nope. Table.

So what's going on? All right, I have totally screwed this up. Well, yeah, I actually messed that up.

I'm sorry. DBA. So Scott, is there no way for you to demonstrate DocuLens?

Yeah, not DocuLens, because the DocuLens container got corrupted. Oh. So unless he can throw that up real quick, then we can show that.

I might be able to. Let me see. While you look into that, we will have a demo at Kim's Informix party, right?

Kim, you can add a mention in the chat if you want about your event soon in Kuala Lumpur. So we will be ready there. Yes.

It's been a mad rush in the last week to get everything corrected here. And actually, the DocuLens demonstrates a little bit better, but... I'm booting that up, and I might be able to show it in a sec.

My server's booting. It's almost ready. We have this, a document analyzer demo in a container, in a Docker container.

So I'm just booting that up. While JC is booting that up. In that first example, and again, we're not going out to the OpenAI, we're adding all these statements into our vector table.

The command I just executed is going to go grab the nearest three of those rows from that vector table. And, since there's only three, we're going to get, again, the three responses here. Okay.

So I'm going to try sharing this. Okay. Let me stop sharing.

As I said, I'm a Zoom newbie, so we'll see how well this goes. Zoom is requesting bypass. Allow.

Screens. I'm going to share my whole screen. Can we see it, JC?

Yep. We see Docker. Okay.

So as I said, I've got a container running with an Informix server and the blade running. And this is just a little demo that one of my colleagues whipped up to use the blade, and I can upload documents using it. So I'm going to upload documents from...

Let's see. I'll try a folder. Presumably, these are documents that are proprietary to your organization.

In my case, I've got some scripts that I wrote and some high-level design specs for features in Informix. And this is just meant to sort of simulate a case where you've got proprietary documents that you don't want out on the web. Oop.

That didn't work. Well looks like my luck is similar to yours, Scott. Let me try one more time.

I'll try an individual document.

Yeah, if you tried an individual, that would work. Or a couple individuals. So right now it's...

Oh. It's possible that my network is not set up correctly. Let's just assume that for the moment.

What this demo did, though, was, it uploads and chunks the documents, vectorizes the chunks, and then you can ask natural language queries against them, and it actually works pretty well. But for right now, I'm not able to upload the documents, probably due to a network issue. So we'll have to give up on that, I'm afraid.

But let's go to... I'm going to stop sharing my screen, and let's go to some Q&A. I've got some questions here that Scott and I will answer.

I see that there's something from John. Yeah, we've got a couple from the actual Q&A window. I'm gonna get back to those in a sec.

So, Scott, can you say how much the Vector Blade costs? That's probably an obvious question people will ask. Right.

So the Vector Blade is free with Informix 15. So there is no cost associated with that. Okay.

And another question is which version of Informix supports the Vector Blade? I've already kind of gone over this. Has to be HL Informix, and it has to be 15.01.

Now, we recommend 15.01.10 or higher, because it's in that version that we programmed auto registration. You can manually register the blade with an earlier version, but, it's just a lot nicer to use the one that does the auto registration. It won't work with older versions like 14.10 or 12.10.

And, so if you're on an older version, that might be a compelling reason to upgrade. So next question, does IBM Informix have this? That one's for you, Scott.

Yes. So IBM Informix does not have the Vector Blade, nor will we allow it to be used within IBM Informix. Okay.

Is there an easy tutorial for me to test out the Vector Blade capabilities? Yes. As I said, when you install the blade, you untar this file into Informix to extend, and there's a subdirectory under there called samples, and that does have four sample scripts that do work, I promise.

Yeah. And they're of increasing complexity and interest. So the first one just registers the blade and does something very rudimentary.

By the time you get to number four, you're doing You're using triggers to insert rows into your Informix table and automatically sync them with your Lance table, and it's more interesting. So- Correct. Next.

Yeah. Just a word on that is, yeah, they all do work because I ran them this morning and last night, and all those were working for me prior to the session here. Yeah.

That's life, I guess. So Scott, for you, how do I get an evaluation license? Right.

So, and you may want to post this in the chat, Nick, but we do have a email address that you can reach out to us to get an evaluation license, and that will come to the product manager, Emily Taylor, and myself, and we will contact you and get the ball rolling on that. Okay. I'm answering some questions here.

We got John asked, "Is there an HCL equivalent to IBM Informix Developer Edition?" So we have these editions on the IBM side, and we don't tend to have the same kind of setup on the HCL side. Yeah. And the question is on that, no, we don't.

We're always the HCL Informix version is the Advanced Enterprise version, comparable to IBM. But HCL only supports and sells that one version. Okay.

Colin asks, "Will the open source dependencies impact ISVs and OEMs that may sell solutions based on the Vector Blade?" This is, I'm afraid, out of my pay grade. Obviously, LanceDB, the C library that we're using is an open-source library, and we have a normal Apache license for that open source library. Scott, or anyone else among the panelists, do you know whether this has an effect on reselling?

I do not. I do not. I do not.

We'll have to get back to Colin. We can take that and do some follow-up. I sure hope not.

But we'll look into that. Yeah, we support the product, okay? So whether we're not going to outsource the support of that if it's part of the product, right?

Correct. Yeah. Yeah.

Okay, I don't see any other questions coming in, so... Oh, we got another one from John. "How do you expect to get new HCL Informix clients without a developer edition, specifically focusing on the Vector Blade, which would seem a big differentiator?" I am going to leave that for Scott.

Yes, I can take that. So, we realize that when we do-- Since we only sell the one version, and every user is different, but we've been very competitive in making sure that we can get you that product. So, that's a discussion that I would be happy to help you with and understand the nuances in your environments.

But, as I've been told several times by my management team, don't let cost prohibit you from making a deal. So... And as Kevin pointed out- That's a serious promise, Scott.

Okay? Oh, yeah. Well, you don't know how much the juggling I do in the background to make everything happen here.

So, Emily has said John Ritson, "It is possible for any interested user to obtain a free trial license for a limited time. But what you get to trial is the HCL Informix Enterprise Edition." So that's kind of cool. Mm-hmm.

But it's a limited time. Correct. "Did you try this data blade in native language?" Oh, so maybe non-English is the question.

That's a really good question. Actually, JGP asked me that question, and I meant to try a French prompt, but I never got around to it. I have not tried that.

It depends on, we're using the OpenAI LLM, and their embedding model, and I'm guessing it's not English only, but I have not tried that. It's a really good question. I'll have to give that a shot.

Okay. When I read the question from Boyko there, I was thinking about when he said native language, I was thinking about Informix 4GL or so. Yeah.

So I would expect that it works with 4GL. So... So JGP, you haven't tried OpenAI embedding model or LLM with French, have you?

No, I have not. And I can't wait to try the blade itself. But I've been a little bit too much on the road lately- Yeah ...

to do that.

But it's very high on my bucket list there. So Boyko, I hope I'm pronouncing your name right. Boyko has mentioned that he wants Bulgarian.

That's going even further. I wouldn't know how to test that, but there's a chance it'll work. As we find out when we use these AI tools, they do things that we don't expect them to do sometimes.

They do them better than we expect. So we'll have to give lots of different languages a shot. So it seems like the questions have wound down here a little bit.

And obviously, we have a few glitches to work out with our demo. But hopefully, you get an idea of what the Blade does, and you're interested in trying it out, and that you've enjoyed our webinar. I guess we can close the webinar now.

Thanks everyone for coming along. Thank you. Bye-bye.

See you again soon. And good-