WEBVTT
00:00:06.240 --> 00:00:08.800
Welcome to Microsoft Community Insights podcast.
00:00:08.880 --> 00:00:13.839
We will share insights from community expert expert to stay up to date with Microsoft episode.
00:00:13.919 --> 00:00:19.039
We will dive into RathRag in Microsoft Agent Framework.
00:00:19.199 --> 00:00:21.280
And today we have a special guest called Christian.
00:00:21.440 --> 00:00:23.039
Do you please introduce yourself, please?
00:00:23.280 --> 00:00:24.879
Yeah, thank you for having me here.
00:00:24.960 --> 00:00:35.039
So, yeah, my name is Christian Klesner, and yeah, I'm quite passionized about AI and agentic development workflows, like you two, right?
00:00:35.200 --> 00:00:36.079
That's why we are here.
00:00:36.240 --> 00:00:43.759
And I'm also now uh for 17 years already a Microsoft MVP, and I saw so many things already with uh with Microsoft.
00:00:44.000 --> 00:00:53.840
But yeah, but I think the AI stuff is really super exciting and it changed so many things, and yeah, really happy to be part of this big transition.
00:00:54.560 --> 00:01:04.000
Okay, so before you before we dive into this episode of the theme, so you be you said you've been an MVP for 17 years.
00:01:04.159 --> 00:01:06.879
Do you want to tell us how you get started?
00:01:07.599 --> 00:01:13.439
Oh in your I started with SharePoint, uh SharePoint 2003.
00:01:14.159 --> 00:01:24.959
So with with.NET One one, I think it was.NET One One, and then you know it came to SharePoint 2003, and this is how I started with that.
00:01:25.040 --> 00:01:26.879
And I actually built some helpers.
00:01:26.959 --> 00:01:29.680
Uh I don't know, I started something.
00:01:29.760 --> 00:01:31.760
It's actually it's a long tradition.
00:01:31.840 --> 00:01:33.359
I do the I love things, you know.
00:01:33.439 --> 00:01:44.640
I had that blog and communities thing, I love SharePoint, which I yeah, which I started and published a lot of examples and web parts, and so I that's suddenly I came an email.
00:01:44.719 --> 00:01:47.840
Oh, I'm an MVP, and I say, Whoa, didn't expect it.
00:01:47.920 --> 00:01:55.040
And yeah, since that I'm in the program, but yeah, recently or many years, I don't do so much SharePoint anymore.
00:01:55.120 --> 00:02:18.400
It's still a it's still a nice technology, but most of the time you explain the people why they have to collaborate, and it's more like and less a tech topic, it's more like uh uh a change management topic to make people working and sharing knowledge, and after I don't know, five, ten years it gets a little bit monotone somehow, and you want to switch.
00:02:18.560 --> 00:02:41.680
Then I also did a lot of mixed reality and then virtual reality as a nice technology, but started then to combine that with AI, but recently it's mostly uh AI that's the highest demand, and I also see you can really make really exciting things with uh with AI in a very short time, and so recently I mostly do AI.
00:02:42.560 --> 00:02:43.599
Okay, yeah.
00:02:43.759 --> 00:02:49.039
So before you were you like a SharePoint MVP, then you switched it to AI.
00:02:49.199 --> 00:02:53.759
So what what was between between uh between mixed reality?
00:02:54.240 --> 00:02:55.120
Oh okay.
00:02:55.280 --> 00:03:00.639
So it's when we now I transition to uh more into that AI topics, yeah.
00:03:00.960 --> 00:03:12.080
Uh okay, it's when you're into mixed reality and then you're you're just curious to learn more and it because it involves it's part of AI, so you're trying to keep up to date with mixed reality.
00:03:12.400 --> 00:03:17.360
Yeah, what I always liked about it, and I I combined the two topics a long time.
00:03:17.520 --> 00:03:21.280
Yeah, it is I always thought AI is a little bit, you know.
00:03:21.599 --> 00:03:28.000
I never liked theory, you know, like hey, taking your phone and then hey, theory and talking with the thing, it doesn't feel natural.
00:03:28.240 --> 00:03:43.680
And what is nice also in in mixed reality or in virtual environments is you can actually give AI a body, you know, it hangs out with you, it has a face, it shows emotion, it feels more natural in just talking with uh with a device.
00:03:43.919 --> 00:03:54.719
But let's see, you know, mixed reality, virtual reality is currently not uh hype topic, and maybe even the robots and you don't have to talk in virtual reality with them.
00:03:54.800 --> 00:04:00.319
If you see in Beijing or something, they're they're running, they're dancing, especially in China.
00:04:00.400 --> 00:04:01.759
You see that a lot, you know.
00:04:02.000 --> 00:04:18.639
So maybe I think in the future we definitely will talk not just with the mobile phone, we will talk with robots uh in virtual reality, whatever it may be more natural, not just talking with yeah, it's uh it's similar to the Microsoft hologram, so in virtual reality.
00:04:18.800 --> 00:04:20.240
So that's an example as well.
00:04:21.040 --> 00:04:26.240
Yeah, I did a lot with with that stuff uh with AI in in uh mixed reality.
00:04:26.480 --> 00:04:28.000
Yeah, we're still doing that actually.
00:04:28.160 --> 00:04:30.240
I'm still doing oh really, yeah.
00:04:30.399 --> 00:04:30.959
Maybe next time.
00:04:31.439 --> 00:04:33.519
Short before publishing a SDK.
00:04:33.680 --> 00:04:39.040
I have a uh by the way, I I did before the iLove SharePoint.
00:04:39.199 --> 00:04:42.240
Now I made a new thing, uh, which started half a year ago.
00:04:42.399 --> 00:04:44.000
ILove Agents AI.
00:04:44.639 --> 00:04:52.879
And uh and part of this, I have already a React wrapper for for the uh foundry voice uh live SDK.
00:04:53.120 --> 00:05:08.800
So that's quite easy to build all kinds of uh voice agents, and I have also variations not published now where you can build agents uh in uh with voice, live voice, uh task in mixed reality in Unity, and they can do more things.
00:05:08.879 --> 00:05:17.040
They can you can set guide points, they can walk with you somewhere, they can look something, they can spawn objects, they can change the environment.
00:05:17.199 --> 00:05:21.439
Um so uh I hope I I will find time to publish that.
00:05:21.519 --> 00:05:24.000
So I'm still doing that stuff, yeah.
00:05:24.319 --> 00:05:31.439
But maybe we maybe we switch soon to the graph rack, if not it's anything, so maybe yeah, it's mixed reality.
00:05:31.680 --> 00:05:33.199
It's quite good mixed reality.
00:05:33.360 --> 00:05:37.360
So okay, so has this episode theme is mixed reality.
00:05:37.759 --> 00:05:39.519
No, graph rags, apologies.
00:05:39.839 --> 00:05:42.319
Uh what is graph frag first?
00:05:42.399 --> 00:05:47.839
The simple question, and why is it used for why is it used for what can you use it for?
00:05:48.319 --> 00:05:52.000
Yeah, wait, I I switch uh here my uh to my screen.
00:05:52.079 --> 00:06:00.879
I will uh what you see there because the the basic thing, what I'm seeing, you see a lot of uh I think you can you can see that, right?
00:06:00.959 --> 00:06:02.959
So yeah, I can see now.
00:06:03.040 --> 00:06:04.079
I'm just okay.
00:06:04.160 --> 00:06:22.480
Yeah, so you I build a lot if you're working with customers, you know, there's currently still well that means still there, there's rec or retrieve augmented results, is a very important topic for customers because you know AI has a lot of power genai.
00:06:22.959 --> 00:06:29.519
But if you talk with chat GPT, whatever, they have only general knowledge, and for sure, this is not the public knowledge.
00:06:29.600 --> 00:06:33.519
Uh the companies have their private knowledge and they don't want to make that public.
00:06:33.600 --> 00:06:56.639
So you have to somehow make the AI aware of your own knowledge, and Rug is the way to uh to do this in in companies, but I see you know, it is not just done if you take all data, you know, you just put it into you throw it into the AI, and then you think, oh, something crazy coming out, like you see on that on that screen here.
00:06:56.720 --> 00:07:00.000
So you still have, especially with RUC, you have the problem.
00:07:00.319 --> 00:07:06.879
It's the oldest kind of IT problem, it's called uh garbage in, garbage out, or even shit in, shit out, like you see here.
00:07:06.959 --> 00:07:20.879
Yeah, and the problem is people think you know, they just put all the stuff, they have all that bad data, all that stuff into the AI, and then the AI makes something super magic out of it, and then finally it becomes great.
00:07:21.040 --> 00:07:41.519
But this uh, like you see here a little bit on the picture where you have uh the image where you have left the bad data, then you have the AI dealing with this, with that shit, and then you see on the on the right side of it, you see, you know, what actually happened is the AI is really good to make not uh true things, sounds very true, you know.
00:07:41.600 --> 00:07:57.920
And so what happened is you have that shit data, but AI really presented you in a really nice way that it even sounds super convincing, and I think it's even more dangerous because it may be completely wrong answers because you already feed it in bad data.
00:07:59.279 --> 00:08:01.920
Yeah, you know you know that problem, right?
00:08:02.240 --> 00:08:05.839
Yeah, so garbage in, garbage out, whatever you put in, you get back out.
00:08:05.920 --> 00:08:13.360
So you have to clean your data, you have to make sure your data is 100% before you get like good results in AI terms.
00:08:13.759 --> 00:08:18.959
Yeah, and then there is another thing is you know what you happen then if you do a Rug solution.
00:08:19.040 --> 00:08:23.920
You have the data and then you chunk them, and then you have small parts, you know.
00:08:24.000 --> 00:08:31.120
There are also a lot of different chunking strategies, how big, how small, semantic chunking, and a lot of different chunk strategies.
00:08:31.199 --> 00:08:56.879
But in the end, you have a lot of chunks, and imagine you have millions of documents, and then you've the AI does a search, uh, a vector search, at least understand that thing, and then it gets bum, a lot of chunks, and it's just fragments of all uh of all the data, so it gets a very how to say fragmented view, and it's hard to see the whole, the big picture because it just gets feeded with a lot of fragments.
00:08:57.200 --> 00:08:58.559
And I call it actually.
00:08:58.639 --> 00:09:04.720
So we I did that session together with my friend uh Said, who uh is actually working now for Neo4j.
00:09:05.120 --> 00:09:12.399
And I call this actually you give the AI, if you use that classic rack approach, I always say you give it junk food.
00:09:14.320 --> 00:09:16.960
Um wait, I have to switch here on the slide.
00:09:17.039 --> 00:09:21.440
And the idea is so then I come with the idea, I call it the knowledge sandwich.
00:09:22.559 --> 00:09:40.000
And it's a little bit like the story is you know, kind of a like on a burger or something, you need a bottom bun somehow, which let's say you have the uh uh the nutrition, the the uh the carbs and things like this, and I say, you know, you have to prepare the data.
00:09:40.159 --> 00:09:53.279
So and I use in many cases if you have really big data, I use a lot of uh Microsoft fabric for this data foundation, and I really prepare the data, and also I did that in my example that I will show later.
00:09:53.440 --> 00:10:11.120
I really make over many stages prepare the data in a good way, and here I also use actually foundry, I use uh content understanding a lot here for actually making sense of this unstructured and semi-structured data and fabric before I feed it into the graph.
00:10:11.519 --> 00:10:29.200
So I do a lot of preparations, and then in the end, foundry, you have a lot of things like foundry IQ, and it is more about and even agent frameworks more about presenting the knowledge, but it's agent framework also doesn't have a tool for uh for ingestion or something like this.
00:10:29.360 --> 00:10:35.120
So I use fabric for that part and more for the visualizing and agentic work on the front.
00:10:35.200 --> 00:10:42.960
I use foundry and I like as kind of a knowledge layer, I really like to use Neo4j as a craft database.
00:10:44.559 --> 00:10:48.720
So what so and what the advantage of the craft database?
00:10:48.879 --> 00:10:57.759
So I have now to switch a little bit to the demo, but I uh I explain it then also uh because yeah, you have to see it somehow.
00:10:57.919 --> 00:10:58.720
What is the difference?
00:10:58.879 --> 00:10:59.200
Yeah.
00:10:59.440 --> 00:11:02.559
Uh where can other people learn more about it?
00:11:02.720 --> 00:11:05.919
It's like is it for your blogging?
00:11:06.000 --> 00:11:07.200
So is does it have a lot of different things?
00:11:07.360 --> 00:11:11.440
Yeah, I have I think I have this contract example in a little lighter version.
00:11:11.519 --> 00:11:15.919
I have that also on GitHub and linked on my I Love Agent AI blog.
00:11:16.159 --> 00:11:24.799
Yeah, and um, I'm also happy I work more because I really like uh I think graph is really uh a big thing with AI.
00:11:24.879 --> 00:11:27.120
It will I it will get much more attention.
00:11:27.279 --> 00:11:46.720
And I'm working recently quite close with Neo4J, and I also will do for Neo4J a lot of specially for the agent framework and for for fountain integration, yeah, uh, will also write the content there or support Neo4J also to uh yeah to get more into the uh Microsoft ecosystem together with my friends.
00:11:47.759 --> 00:12:00.159
Yeah, so for those who want to learn more about Graphrag, I would uh I would attach the the his blog or details on the show notes so you can get to if you want details.
00:12:00.879 --> 00:12:16.399
And there will also be much more uh there is already on the Neo for J they have also a lab site, and there's also how to integrate it with uh with with Foundry and a lot of uh labs, which I will improve to, especially for agent framework.
00:12:16.879 --> 00:12:17.919
Nice, brilliant.
00:12:18.000 --> 00:12:24.799
Yeah, it was it's quite good how you can visualize that in some before that you can actually visualize the data.
00:12:24.960 --> 00:12:34.159
So that's so the bigger the the bigger the data on a query, the larger the visualization, the graph would be, you know.
00:12:35.039 --> 00:12:38.480
Yeah, so it is it's really a big difference from the quality.
00:12:38.639 --> 00:12:43.120
If if the I really has it, it's hard to do that later.
00:12:43.440 --> 00:12:48.799
But the you know, craft is not the craft data, but it's not a new thing.
00:12:48.960 --> 00:12:54.159
It was always good, and actually writing the cipher career, which if you have a good craft, it's not so hard.
00:12:54.240 --> 00:13:01.759
Yeah, the problem all people usually have, and this is also where I help a lot, is actually building a craft.
00:13:01.840 --> 00:13:19.360
This is actually the hardest part because you have all that non-craft data lying there, you know, in in I don't know, terabytes of data, yeah, and it's there is you know it did that's what I tried at the beginning, and this is how I did it actually in my uh simpler blog post.
00:13:19.679 --> 00:13:31.360
And you use the LLM, does extraction, but not all of that is right, and then you put it into the graph, and then the graph also has wrong data, and then the AI will also present wrong data.
00:13:31.759 --> 00:13:50.480
But this is not a special problem of craft, this is the general shit-in, shit out thing, and then I started to use fabric, and I make a lot of more data clean up and and different things, and then I could increase the quality massively, but it was really a longer pipe, and it's I think it's a lot underestimated.
00:13:50.639 --> 00:13:55.120
Yeah, how much you have to care about that ingestion part.
00:13:55.200 --> 00:13:59.279
This is and AI also helps there because you know, content understanding.
00:13:59.360 --> 00:14:02.720
I use a lot of content understanding, I'm actually quite happy with it.
00:14:02.879 --> 00:14:15.120
You know, it has to PDF cracking, this is not so hard, but then you can also have this uh you know custom schemas where you can extract, like you extract the contact clauses, and it isn't content understanding.
00:14:15.200 --> 00:14:23.600
If you do custom analyzer, you kind of write for everything like a mini prompt uh for for your different fields that it should extract.
00:14:24.000 --> 00:14:35.279
So would you recommend so you would you recommend someone to go use uh max fabric instead of using graph first because of the data?
00:14:35.919 --> 00:14:48.799
If you have really big data, and I have situations with uh with customers that you really really big data, then fabric is a good way to do this because you can do experiments with you know with Python.
00:14:49.039 --> 00:14:54.720
Um and but later on it scales really on many nodes for for bigger parsing.
00:14:55.039 --> 00:14:58.080
Um I think it depends a little bit.
00:14:58.159 --> 00:15:06.080
I I like to use it, um what I really always use meanwhile is I think content understanding got quite good.
00:15:06.399 --> 00:15:09.919
Of course, it does also for large PDFs, the paging and everything.
00:15:10.080 --> 00:15:18.240
So I would definitely use content understanding as part of the of the ingestion layer for semi and unstructured data.
00:15:18.399 --> 00:15:21.600
So and fabric depends a little bit on the data.
00:15:21.840 --> 00:15:57.600
What I didn't like and fabric too much is currently you're so productive with you know with all that agentic tools like uh you know, like GitHub copilot, like uh clawed code and things like this, and it works also with the GitHub copilot plugin, but it's not works not so good as if you just do it plain, and the productivity is a little bit lower if you do this instead of yeah, because the tooling, uh the agentic tooling is better currently without fabric, but Microsoft's really hard working to uh you know to improve that agentic tool.
00:15:58.960 --> 00:16:02.559
Because nobody wants to write big SQL statements or all that stuff.
00:16:02.720 --> 00:16:03.519
Uh no.
00:16:05.919 --> 00:16:10.799
Maybe they probably will put an MCP in Fabric one time.
00:16:10.879 --> 00:16:12.240
So hopefully.
00:16:12.720 --> 00:16:19.600
Yeah, they have found we uh they have uh fabric IQ, they have already this that stuff.
00:16:19.759 --> 00:16:36.559
And I just saw today they also have some new AI tools for data wrangler, and for sure everything is AI, and maybe my my last thing is really, you know, I really like developing, but I also didn't show now the the agent framework code anymore.
00:16:36.879 --> 00:16:54.000
Because to be honest, you know, I think just writing code yourself is nice, some maybe for a little bit learning, but in reality, if you do if you're really good with agentic tooling, it's much more productive than you write anything yourself.
00:16:54.080 --> 00:17:06.799
The AI is really good, you just have to teach it, and so I think we are on a way from you know, from developer to an agentic engineer, and I really I think I barely write code, even though I like that a lot.
00:17:06.880 --> 00:17:08.880
But the productivity with AI is so much.
00:17:10.640 --> 00:17:17.359
It's the same, like we're like reviewers, we look at we review our code and stuff instead of writing it.
00:17:17.440 --> 00:17:32.000
So we get the AI to write it, we just look for the our code for like any like if it's good quality, like proofreading and setting a PRs and stuff, good quality in terms of the problem that I see there is it is for somebody, you know.
00:17:32.240 --> 00:17:40.319
You see, I'm already a long time in that business, even like you know, um to over 20 years and doing stuff like this already.
00:17:40.480 --> 00:18:06.880
And you know, I started with HTML uh self-html, I don't know, 5-0 or something, you know, just the beginning of JavaScript and HTML, and you know all that technology stack all behind it, and then you become kind of an architect that knows Azure, Microsoft technology, and you have to know all of the things, and and if you do you see if you monitor the AI and they get crazy, they do something.
00:18:06.960 --> 00:18:10.400
Wow, you think, oh, what the hell did you do there?
00:18:10.640 --> 00:18:13.200
And why do you just do that like this?
00:18:13.359 --> 00:18:19.279
It's like a it's somehow like a junior developer with a PhD, very high ability.
00:18:19.599 --> 00:18:28.640
Yeah, you have no idea, and then it hacks there, and then you say, ah that's what that's what giving instruction is very important to the AI.
00:18:28.960 --> 00:18:36.799
But if you know if you know all of that thing and you can control it like a team of junior developers, it's really awesome.
00:18:37.039 --> 00:18:52.000
But if if you if people now start developing with this and they don't have all that background about entra and how all of that things, authentication, security, protocols working, and then they're just starting with AI, this is kind of even dangerous.
00:18:52.079 --> 00:18:54.480
They produce results, it looks nice, yeah.
00:18:54.880 --> 00:18:59.759
But behind this is horrible and unsecure and dangerous.
00:18:59.920 --> 00:19:17.200
So for people really knowing what they're doing and a good architect, it's wow, it's magic, but it's also dangerous for uh if you are not so yeah, because you trust the AI, and it there's a lot of mistakes, bad architectural decisions, bad security decisions.
00:19:17.599 --> 00:19:28.240
Yeah, so you never ever trust an agent, you have to trust the human behind the agents, that great agent, because we're the one that behind engineer the agent itself.
00:19:28.559 --> 00:19:45.279
You become more to to a reviewer to writing plans and later uh do the reviews and and guide this, but the really you know, writing the the the single quote lines, this uh I think that so long anymore.
00:19:45.519 --> 00:19:50.799
Yeah, we will be just governing, so we'll put in the guardrails for the agent evaluation does.
00:19:51.039 --> 00:19:55.200
Yeah, so yeah, it will be interesting in a few years.
00:19:55.279 --> 00:19:57.039
Now, next year is gonna be interesting.
00:19:57.279 --> 00:20:04.000
So so as this episode is coming to an end, uh we just love to find out about more about our guests.
00:20:04.079 --> 00:20:06.799
So, Kristen, are you going to any other events?
00:20:06.960 --> 00:20:10.160
Any other like virtual reality uh talking?
00:20:10.480 --> 00:20:14.720
Are you talking to any about graph rack daily in any events?
00:20:14.880 --> 00:20:29.519
Uh uh yeah, my next events that I actually doing is uh fly, I think in two weeks to Bangkok to the agent camp there, and directly from there I do inside together, uh the Agent Con in Berlin.
00:20:30.400 --> 00:20:44.319
So I fly directly from there to the Agent Con Berlin, which I'm actually working to inside, and then uh I think one month later in June, we go to Brazil, so I come a little bit around to the Agent Con uh in Rio.
00:20:44.799 --> 00:20:50.079
Yeah, so you have to pack for all temperature, you go around the world.
00:20:50.400 --> 00:20:52.240
It's yeah, it's all fun anyway.
00:20:52.480 --> 00:21:02.880
So I I think this personal relationships getting with AI, you know, when AI doing more of that logical work, I think actually, what should we do?
00:21:02.960 --> 00:21:03.680
What should you do?
00:21:03.759 --> 00:21:09.519
I think that emotional thing or having connections and just hang out and have a drink together.
00:21:09.599 --> 00:21:18.799
I think it even gets more important in the times of AI because that somehow sets us apart from uh from the AI machines.
00:21:19.759 --> 00:21:25.519
We need uh human connections, we need the actual not the agent itself.
00:21:25.599 --> 00:21:34.160
Don't get me wrong, agents is amazing, but we need the actual one-to-one connection than agent because we're behind, we don't want to create an agent at the at the end.
00:21:34.480 --> 00:21:39.599
So we let the agents do the work and we you know hanging out at the beach and have some drinks, Niklas, huh?
00:21:39.680 --> 00:21:40.559
That's the plan.
00:21:41.119 --> 00:21:43.759
Uh yeah, but you still have to look after the agent.
00:21:43.839 --> 00:21:52.559
You have to do the ground governance at the agents, you have to do the security like behind the scene, like uh like the infrastructure basically.
00:21:53.039 --> 00:21:59.839
Uh so yeah, so if you want to watch his talk at uh next month for agent camp bankrupt.
00:22:00.079 --> 00:22:03.279
Go to Bangkok and go to Thailand.
00:22:03.359 --> 00:22:04.079
It's an amazing place.
00:22:04.480 --> 00:22:04.960
Kristen said.
00:22:05.039 --> 00:22:06.640
I've been to there once and it's quite good.
00:22:06.880 --> 00:22:11.200
You get uh you get there's always food every night, 24 hours.
00:22:11.279 --> 00:22:13.519
So it's amazing street food as well.
00:22:13.599 --> 00:22:16.480
So if you like Asian food, it's amazing there as well.
00:22:16.559 --> 00:22:19.359
It's very cheap as well, so you can do that as well.
00:22:19.920 --> 00:22:32.559
Yeah, and hope you hope everyone learned a bit about the importance of using crag rag and then when to use it and how to when to use fabric when you have to clean your data more.
00:22:32.720 --> 00:22:34.640
So thanks a lot, Kristen.
00:22:35.519 --> 00:22:36.000
Bye.
00:22:36.319 --> 00:22:36.880
Bye.
00:00:06.240 --> 00:00:08.800
Welcome to Microsoft Community Insights podcast.
00:00:08.880 --> 00:00:13.839
We will share insights from community expert expert to stay up to date with Microsoft episode.
00:00:13.919 --> 00:00:19.039
We will dive into RathRag in Microsoft Agent Framework.
00:00:19.199 --> 00:00:21.280
And today we have a special guest called Christian.
00:00:21.440 --> 00:00:23.039
Do you please introduce yourself, please?
00:00:23.280 --> 00:00:24.879
Yeah, thank you for having me here.
00:00:24.960 --> 00:00:35.039
So, yeah, my name is Christian Klesner, and yeah, I'm quite passionized about AI and agentic development workflows, like you two, right?
00:00:35.200 --> 00:00:36.079
That's why we are here.
00:00:36.240 --> 00:00:43.759
And I'm also now uh for 17 years already a Microsoft MVP, and I saw so many things already with uh with Microsoft.
00:00:44.000 --> 00:00:53.840
But yeah, but I think the AI stuff is really super exciting and it changed so many things, and yeah, really happy to be part of this big transition.
00:00:54.560 --> 00:01:04.000
Okay, so before you before we dive into this episode of the theme, so you be you said you've been an MVP for 17 years.
00:01:04.159 --> 00:01:06.879
Do you want to tell us how you get started?
00:01:07.599 --> 00:01:13.439
Oh in your I started with SharePoint, uh SharePoint 2003.
00:01:14.159 --> 00:01:24.959
So with with.NET One one, I think it was.NET One One, and then you know it came to SharePoint 2003, and this is how I started with that.
00:01:25.040 --> 00:01:26.879
And I actually built some helpers.
00:01:26.959 --> 00:01:29.680
Uh I don't know, I started something.
00:01:29.760 --> 00:01:31.760
It's actually it's a long tradition.
00:01:31.840 --> 00:01:33.359
I do the I love things, you know.
00:01:33.439 --> 00:01:44.640
I had that blog and communities thing, I love SharePoint, which I yeah, which I started and published a lot of examples and web parts, and so I that's suddenly I came an email.
00:01:44.719 --> 00:01:47.840
Oh, I'm an MVP, and I say, Whoa, didn't expect it.
00:01:47.920 --> 00:01:55.040
And yeah, since that I'm in the program, but yeah, recently or many years, I don't do so much SharePoint anymore.
00:01:55.120 --> 00:02:18.400
It's still a it's still a nice technology, but most of the time you explain the people why they have to collaborate, and it's more like and less a tech topic, it's more like uh uh a change management topic to make people working and sharing knowledge, and after I don't know, five, ten years it gets a little bit monotone somehow, and you want to switch.
00:02:18.560 --> 00:02:41.680
Then I also did a lot of mixed reality and then virtual reality as a nice technology, but started then to combine that with AI, but recently it's mostly uh AI that's the highest demand, and I also see you can really make really exciting things with uh with AI in a very short time, and so recently I mostly do AI.
00:02:42.560 --> 00:02:43.599
Okay, yeah.
00:02:43.759 --> 00:02:49.039
So before you were you like a SharePoint MVP, then you switched it to AI.
00:02:49.199 --> 00:02:53.759
So what what was between between uh between mixed reality?
00:02:54.240 --> 00:02:55.120
Oh okay.
00:02:55.280 --> 00:03:00.639
So it's when we now I transition to uh more into that AI topics, yeah.
00:03:00.960 --> 00:03:12.080
Uh okay, it's when you're into mixed reality and then you're you're just curious to learn more and it because it involves it's part of AI, so you're trying to keep up to date with mixed reality.
00:03:12.400 --> 00:03:17.360
Yeah, what I always liked about it, and I I combined the two topics a long time.
00:03:17.520 --> 00:03:21.280
Yeah, it is I always thought AI is a little bit, you know.
00:03:21.599 --> 00:03:28.000
I never liked theory, you know, like hey, taking your phone and then hey, theory and talking with the thing, it doesn't feel natural.
00:03:28.240 --> 00:03:43.680
And what is nice also in in mixed reality or in virtual environments is you can actually give AI a body, you know, it hangs out with you, it has a face, it shows emotion, it feels more natural in just talking with uh with a device.
00:03:43.919 --> 00:03:54.719
But let's see, you know, mixed reality, virtual reality is currently not uh hype topic, and maybe even the robots and you don't have to talk in virtual reality with them.
00:03:54.800 --> 00:04:00.319
If you see in Beijing or something, they're they're running, they're dancing, especially in China.
00:04:00.400 --> 00:04:01.759
You see that a lot, you know.
00:04:02.000 --> 00:04:18.639
So maybe I think in the future we definitely will talk not just with the mobile phone, we will talk with robots uh in virtual reality, whatever it may be more natural, not just talking with yeah, it's uh it's similar to the Microsoft hologram, so in virtual reality.
00:04:18.800 --> 00:04:20.240
So that's an example as well.
00:04:21.040 --> 00:04:26.240
Yeah, I did a lot with with that stuff uh with AI in in uh mixed reality.
00:04:26.480 --> 00:04:28.000
Yeah, we're still doing that actually.
00:04:28.160 --> 00:04:30.240
I'm still doing oh really, yeah.
00:04:30.399 --> 00:04:30.959
Maybe next time.
00:04:31.439 --> 00:04:33.519
Short before publishing a SDK.
00:04:33.680 --> 00:04:39.040
I have a uh by the way, I I did before the iLove SharePoint.
00:04:39.199 --> 00:04:42.240
Now I made a new thing, uh, which started half a year ago.
00:04:42.399 --> 00:04:44.000
ILove Agents AI.
00:04:44.639 --> 00:04:52.879
And uh and part of this, I have already a React wrapper for for the uh foundry voice uh live SDK.
00:04:53.120 --> 00:05:08.800
So that's quite easy to build all kinds of uh voice agents, and I have also variations not published now where you can build agents uh in uh with voice, live voice, uh task in mixed reality in Unity, and they can do more things.
00:05:08.879 --> 00:05:17.040
They can you can set guide points, they can walk with you somewhere, they can look something, they can spawn objects, they can change the environment.
00:05:17.199 --> 00:05:21.439
Um so uh I hope I I will find time to publish that.
00:05:21.519 --> 00:05:24.000
So I'm still doing that stuff, yeah.
00:05:24.319 --> 00:05:31.439
But maybe we maybe we switch soon to the graph rack, if not it's anything, so maybe yeah, it's mixed reality.
00:05:31.680 --> 00:05:33.199
It's quite good mixed reality.
00:05:33.360 --> 00:05:37.360
So okay, so has this episode theme is mixed reality.
00:05:37.759 --> 00:05:39.519
No, graph rags, apologies.
00:05:39.839 --> 00:05:42.319
Uh what is graph frag first?
00:05:42.399 --> 00:05:47.839
The simple question, and why is it used for why is it used for what can you use it for?
00:05:48.319 --> 00:05:52.000
Yeah, wait, I I switch uh here my uh to my screen.
00:05:52.079 --> 00:06:00.879
I will uh what you see there because the the basic thing, what I'm seeing, you see a lot of uh I think you can you can see that, right?
00:06:00.959 --> 00:06:02.959
So yeah, I can see now.
00:06:03.040 --> 00:06:04.079
I'm just okay.
00:06:04.160 --> 00:06:22.480
Yeah, so you I build a lot if you're working with customers, you know, there's currently still well that means still there, there's rec or retrieve augmented results, is a very important topic for customers because you know AI has a lot of power genai.
00:06:22.959 --> 00:06:29.519
But if you talk with chat GPT, whatever, they have only general knowledge, and for sure, this is not the public knowledge.
00:06:29.600 --> 00:06:33.519
Uh the companies have their private knowledge and they don't want to make that public.
00:06:33.600 --> 00:06:56.639
So you have to somehow make the AI aware of your own knowledge, and Rug is the way to uh to do this in in companies, but I see you know, it is not just done if you take all data, you know, you just put it into you throw it into the AI, and then you think, oh, something crazy coming out, like you see on that on that screen here.
00:06:56.720 --> 00:07:00.000
So you still have, especially with RUC, you have the problem.
00:07:00.319 --> 00:07:06.879
It's the oldest kind of IT problem, it's called uh garbage in, garbage out, or even shit in, shit out, like you see here.
00:07:06.959 --> 00:07:20.879
Yeah, and the problem is people think you know, they just put all the stuff, they have all that bad data, all that stuff into the AI, and then the AI makes something super magic out of it, and then finally it becomes great.
00:07:21.040 --> 00:07:41.519
But this uh, like you see here a little bit on the picture where you have uh the image where you have left the bad data, then you have the AI dealing with this, with that shit, and then you see on the on the right side of it, you see, you know, what actually happened is the AI is really good to make not uh true things, sounds very true, you know.
00:07:41.600 --> 00:07:57.920
And so what happened is you have that shit data, but AI really presented you in a really nice way that it even sounds super convincing, and I think it's even more dangerous because it may be completely wrong answers because you already feed it in bad data.
00:07:59.279 --> 00:08:01.920
Yeah, you know you know that problem, right?
00:08:02.240 --> 00:08:05.839
Yeah, so garbage in, garbage out, whatever you put in, you get back out.
00:08:05.920 --> 00:08:13.360
So you have to clean your data, you have to make sure your data is 100% before you get like good results in AI terms.
00:08:13.759 --> 00:08:18.959
Yeah, and then there is another thing is you know what you happen then if you do a Rug solution.
00:08:19.040 --> 00:08:23.920
You have the data and then you chunk them, and then you have small parts, you know.
00:08:24.000 --> 00:08:31.120
There are also a lot of different chunking strategies, how big, how small, semantic chunking, and a lot of different chunk strategies.
00:08:31.199 --> 00:08:56.879
But in the end, you have a lot of chunks, and imagine you have millions of documents, and then you've the AI does a search, uh, a vector search, at least understand that thing, and then it gets bum, a lot of chunks, and it's just fragments of all uh of all the data, so it gets a very how to say fragmented view, and it's hard to see the whole, the big picture because it just gets feeded with a lot of fragments.
00:08:57.200 --> 00:08:58.559
And I call it actually.
00:08:58.639 --> 00:09:04.720
So we I did that session together with my friend uh Said, who uh is actually working now for Neo4j.
00:09:05.120 --> 00:09:12.399
And I call this actually you give the AI, if you use that classic rack approach, I always say you give it junk food.
00:09:14.320 --> 00:09:16.960
Um wait, I have to switch here on the slide.
00:09:17.039 --> 00:09:21.440
And the idea is so then I come with the idea, I call it the knowledge sandwich.
00:09:22.559 --> 00:09:40.000
And it's a little bit like the story is you know, kind of a like on a burger or something, you need a bottom bun somehow, which let's say you have the uh uh the nutrition, the the uh the carbs and things like this, and I say, you know, you have to prepare the data.
00:09:40.159 --> 00:09:53.279
So and I use in many cases if you have really big data, I use a lot of uh Microsoft fabric for this data foundation, and I really prepare the data, and also I did that in my example that I will show later.
00:09:53.440 --> 00:10:11.120
I really make over many stages prepare the data in a good way, and here I also use actually foundry, I use uh content understanding a lot here for actually making sense of this unstructured and semi-structured data and fabric before I feed it into the graph.
00:10:11.519 --> 00:10:29.200
So I do a lot of preparations, and then in the end, foundry, you have a lot of things like foundry IQ, and it is more about and even agent frameworks more about presenting the knowledge, but it's agent framework also doesn't have a tool for uh for ingestion or something like this.
00:10:29.360 --> 00:10:35.120
So I use fabric for that part and more for the visualizing and agentic work on the front.
00:10:35.200 --> 00:10:42.960
I use foundry and I like as kind of a knowledge layer, I really like to use Neo4j as a craft database.
00:10:44.559 --> 00:10:48.720
So what so and what the advantage of the craft database?
00:10:48.879 --> 00:10:57.759
So I have now to switch a little bit to the demo, but I uh I explain it then also uh because yeah, you have to see it somehow.
00:10:57.919 --> 00:10:58.720
What is the difference?
00:10:58.879 --> 00:10:59.200
Yeah.
00:10:59.440 --> 00:11:02.559
Uh where can other people learn more about it?
00:11:02.720 --> 00:11:05.919
It's like is it for your blogging?
00:11:06.000 --> 00:11:07.200
So is does it have a lot of different things?
00:11:07.360 --> 00:11:11.440
Yeah, I have I think I have this contract example in a little lighter version.
00:11:11.519 --> 00:11:15.919
I have that also on GitHub and linked on my I Love Agent AI blog.
00:11:16.159 --> 00:11:24.799
Yeah, and um, I'm also happy I work more because I really like uh I think graph is really uh a big thing with AI.
00:11:24.879 --> 00:11:27.120
It will I it will get much more attention.
00:11:27.279 --> 00:11:46.720
And I'm working recently quite close with Neo4J, and I also will do for Neo4J a lot of specially for the agent framework and for for fountain integration, yeah, uh, will also write the content there or support Neo4J also to uh yeah to get more into the uh Microsoft ecosystem together with my friends.
00:11:47.759 --> 00:12:00.159
Yeah, so for those who want to learn more about Graphrag, I would uh I would attach the the his blog or details on the show notes so you can get to if you want details.
00:12:00.879 --> 00:12:16.399
And there will also be much more uh there is already on the Neo for J they have also a lab site, and there's also how to integrate it with uh with with Foundry and a lot of uh labs, which I will improve to, especially for agent framework.
00:12:16.879 --> 00:12:17.919
Nice, brilliant.
00:12:18.000 --> 00:12:24.799
Yeah, it was it's quite good how you can visualize that in some before that you can actually visualize the data.
00:12:24.960 --> 00:12:34.159
So that's so the bigger the the bigger the data on a query, the larger the visualization, the graph would be, you know.
00:12:35.039 --> 00:12:38.480
Yeah, so it is it's really a big difference from the quality.
00:12:38.639 --> 00:12:43.120
If if the I really has it, it's hard to do that later.
00:12:43.440 --> 00:12:48.799
But the you know, craft is not the craft data, but it's not a new thing.
00:12:48.960 --> 00:12:54.159
It was always good, and actually writing the cipher career, which if you have a good craft, it's not so hard.
00:12:54.240 --> 00:13:01.759
Yeah, the problem all people usually have, and this is also where I help a lot, is actually building a craft.
00:13:01.840 --> 00:13:19.360
This is actually the hardest part because you have all that non-craft data lying there, you know, in in I don't know, terabytes of data, yeah, and it's there is you know it did that's what I tried at the beginning, and this is how I did it actually in my uh simpler blog post.
00:13:19.679 --> 00:13:31.360
And you use the LLM, does extraction, but not all of that is right, and then you put it into the graph, and then the graph also has wrong data, and then the AI will also present wrong data.
00:13:31.759 --> 00:13:50.480
But this is not a special problem of craft, this is the general shit-in, shit out thing, and then I started to use fabric, and I make a lot of more data clean up and and different things, and then I could increase the quality massively, but it was really a longer pipe, and it's I think it's a lot underestimated.
00:13:50.639 --> 00:13:55.120
Yeah, how much you have to care about that ingestion part.
00:13:55.200 --> 00:13:59.279
This is and AI also helps there because you know, content understanding.
00:13:59.360 --> 00:14:02.720
I use a lot of content understanding, I'm actually quite happy with it.
00:14:02.879 --> 00:14:15.120
You know, it has to PDF cracking, this is not so hard, but then you can also have this uh you know custom schemas where you can extract, like you extract the contact clauses, and it isn't content understanding.
00:14:15.200 --> 00:14:23.600
If you do custom analyzer, you kind of write for everything like a mini prompt uh for for your different fields that it should extract.
00:14:24.000 --> 00:14:35.279
So would you recommend so you would you recommend someone to go use uh max fabric instead of using graph first because of the data?
00:14:35.919 --> 00:14:48.799
If you have really big data, and I have situations with uh with customers that you really really big data, then fabric is a good way to do this because you can do experiments with you know with Python.
00:14:49.039 --> 00:14:54.720
Um and but later on it scales really on many nodes for for bigger parsing.
00:14:55.039 --> 00:14:58.080
Um I think it depends a little bit.
00:14:58.159 --> 00:15:06.080
I I like to use it, um what I really always use meanwhile is I think content understanding got quite good.
00:15:06.399 --> 00:15:09.919
Of course, it does also for large PDFs, the paging and everything.
00:15:10.080 --> 00:15:18.240
So I would definitely use content understanding as part of the of the ingestion layer for semi and unstructured data.
00:15:18.399 --> 00:15:21.600
So and fabric depends a little bit on the data.
00:15:21.840 --> 00:15:57.600
What I didn't like and fabric too much is currently you're so productive with you know with all that agentic tools like uh you know, like GitHub copilot, like uh clawed code and things like this, and it works also with the GitHub copilot plugin, but it's not works not so good as if you just do it plain, and the productivity is a little bit lower if you do this instead of yeah, because the tooling, uh the agentic tooling is better currently without fabric, but Microsoft's really hard working to uh you know to improve that agentic tool.
00:15:58.960 --> 00:16:02.559
Because nobody wants to write big SQL statements or all that stuff.
00:16:02.720 --> 00:16:03.519
Uh no.
00:16:05.919 --> 00:16:10.799
Maybe they probably will put an MCP in Fabric one time.
00:16:10.879 --> 00:16:12.240
So hopefully.
00:16:12.720 --> 00:16:19.600
Yeah, they have found we uh they have uh fabric IQ, they have already this that stuff.
00:16:19.759 --> 00:16:36.559
And I just saw today they also have some new AI tools for data wrangler, and for sure everything is AI, and maybe my my last thing is really, you know, I really like developing, but I also didn't show now the the agent framework code anymore.
00:16:36.879 --> 00:16:54.000
Because to be honest, you know, I think just writing code yourself is nice, some maybe for a little bit learning, but in reality, if you do if you're really good with agentic tooling, it's much more productive than you write anything yourself.
00:16:54.080 --> 00:17:06.799
The AI is really good, you just have to teach it, and so I think we are on a way from you know, from developer to an agentic engineer, and I really I think I barely write code, even though I like that a lot.
00:17:06.880 --> 00:17:08.880
But the productivity with AI is so much.
00:17:10.640 --> 00:17:17.359
It's the same, like we're like reviewers, we look at we review our code and stuff instead of writing it.
00:17:17.440 --> 00:17:32.000
So we get the AI to write it, we just look for the our code for like any like if it's good quality, like proofreading and setting a PRs and stuff, good quality in terms of the problem that I see there is it is for somebody, you know.
00:17:32.240 --> 00:17:40.319
You see, I'm already a long time in that business, even like you know, um to over 20 years and doing stuff like this already.
00:17:40.480 --> 00:18:06.880
And you know, I started with HTML uh self-html, I don't know, 5-0 or something, you know, just the beginning of JavaScript and HTML, and you know all that technology stack all behind it, and then you become kind of an architect that knows Azure, Microsoft technology, and you have to know all of the things, and and if you do you see if you monitor the AI and they get crazy, they do something.
00:18:06.960 --> 00:18:10.400
Wow, you think, oh, what the hell did you do there?
00:18:10.640 --> 00:18:13.200
And why do you just do that like this?
00:18:13.359 --> 00:18:19.279
It's like a it's somehow like a junior developer with a PhD, very high ability.
00:18:19.599 --> 00:18:28.640
Yeah, you have no idea, and then it hacks there, and then you say, ah that's what that's what giving instruction is very important to the AI.
00:18:28.960 --> 00:18:36.799
But if you know if you know all of that thing and you can control it like a team of junior developers, it's really awesome.
00:18:37.039 --> 00:18:52.000
But if if you if people now start developing with this and they don't have all that background about entra and how all of that things, authentication, security, protocols working, and then they're just starting with AI, this is kind of even dangerous.
00:18:52.079 --> 00:18:54.480
They produce results, it looks nice, yeah.
00:18:54.880 --> 00:18:59.759
But behind this is horrible and unsecure and dangerous.
00:18:59.920 --> 00:19:17.200
So for people really knowing what they're doing and a good architect, it's wow, it's magic, but it's also dangerous for uh if you are not so yeah, because you trust the AI, and it there's a lot of mistakes, bad architectural decisions, bad security decisions.
00:19:17.599 --> 00:19:28.240
Yeah, so you never ever trust an agent, you have to trust the human behind the agents, that great agent, because we're the one that behind engineer the agent itself.
00:19:28.559 --> 00:19:45.279
You become more to to a reviewer to writing plans and later uh do the reviews and and guide this, but the really you know, writing the the the single quote lines, this uh I think that so long anymore.
00:19:45.519 --> 00:19:50.799
Yeah, we will be just governing, so we'll put in the guardrails for the agent evaluation does.
00:19:51.039 --> 00:19:55.200
Yeah, so yeah, it will be interesting in a few years.
00:19:55.279 --> 00:19:57.039
Now, next year is gonna be interesting.
00:19:57.279 --> 00:20:04.000
So so as this episode is coming to an end, uh we just love to find out about more about our guests.
00:20:04.079 --> 00:20:06.799
So, Kristen, are you going to any other events?
00:20:06.960 --> 00:20:10.160
Any other like virtual reality uh talking?
00:20:10.480 --> 00:20:14.720
Are you talking to any about graph rack daily in any events?
00:20:14.880 --> 00:20:29.519
Uh uh yeah, my next events that I actually doing is uh fly, I think in two weeks to Bangkok to the agent camp there, and directly from there I do inside together, uh the Agent Con in Berlin.
00:20:30.400 --> 00:20:44.319
So I fly directly from there to the Agent Con Berlin, which I'm actually working to inside, and then uh I think one month later in June, we go to Brazil, so I come a little bit around to the Agent Con uh in Rio.
00:20:44.799 --> 00:20:50.079
Yeah, so you have to pack for all temperature, you go around the world.
00:20:50.400 --> 00:20:52.240
It's yeah, it's all fun anyway.
00:20:52.480 --> 00:21:02.880
So I I think this personal relationships getting with AI, you know, when AI doing more of that logical work, I think actually, what should we do?
00:21:02.960 --> 00:21:03.680
What should you do?
00:21:03.759 --> 00:21:09.519
I think that emotional thing or having connections and just hang out and have a drink together.
00:21:09.599 --> 00:21:18.799
I think it even gets more important in the times of AI because that somehow sets us apart from uh from the AI machines.
00:21:19.759 --> 00:21:25.519
We need uh human connections, we need the actual not the agent itself.
00:21:25.599 --> 00:21:34.160
Don't get me wrong, agents is amazing, but we need the actual one-to-one connection than agent because we're behind, we don't want to create an agent at the at the end.
00:21:34.480 --> 00:21:39.599
So we let the agents do the work and we you know hanging out at the beach and have some drinks, Niklas, huh?
00:21:39.680 --> 00:21:40.559
That's the plan.
00:21:41.119 --> 00:21:43.759
Uh yeah, but you still have to look after the agent.
00:21:43.839 --> 00:21:52.559
You have to do the ground governance at the agents, you have to do the security like behind the scene, like uh like the infrastructure basically.
00:21:53.039 --> 00:21:59.839
Uh so yeah, so if you want to watch his talk at uh next month for agent camp bankrupt.
00:22:00.079 --> 00:22:03.279
Go to Bangkok and go to Thailand.
00:22:03.359 --> 00:22:04.079
It's an amazing place.
00:22:04.480 --> 00:22:04.960
Kristen said.
00:22:05.039 --> 00:22:06.640
I've been to there once and it's quite good.
00:22:06.880 --> 00:22:11.200
You get uh you get there's always food every night, 24 hours.
00:22:11.279 --> 00:22:13.519
So it's amazing street food as well.
00:22:13.599 --> 00:22:16.480
So if you like Asian food, it's amazing there as well.
00:22:16.559 --> 00:22:19.359
It's very cheap as well, so you can do that as well.
00:22:19.920 --> 00:22:32.559
Yeah, and hope you hope everyone learned a bit about the importance of using crag rag and then when to use it and how to when to use fabric when you have to clean your data more.
00:22:32.720 --> 00:22:34.640
So thanks a lot, Kristen.
00:22:35.519 --> 00:22:36.000
Bye.
00:22:36.319 --> 00:22:36.880
Bye.
