SharePoint to Lakehouse with Lakeflow Connect — Top 5 Powerful GenAI Patterns

Summary

The Lakeflow Connect SharePoint connector ingests both structured data—Excel files, CSVs, and lists—and unstructured content including PDFs, Office documents, and images into the Databricks lakehouse as streaming tables enriched with SharePoint metadata, all governed through Unity Catalog.
Five low-lift, high-impact GenAI patterns are demonstrated: OCR for historical scanned TIFF documents, AI-powered BI dashboards from Excel and list data, AI summarization of PDF reports, natural language querying of PowerPoint decks in a Genie Room, and consolidation of many CSV reports into a single streaming table.
All ingestion pipelines run incrementally and are governed through Unity Catalog, preserving SharePoint data lineage and enabling downstream analytics, search, and AI workloads from a single unified lakehouse foundation.

SharePoint to Lakehouse with Lakeflow Connect — Top 5 Powerful GenAI Patterns

Many organizations have rich SharePoint data but struggle to use it at scale for intelligence, reporting, and automation. In this demo-heavy session, we demonstrate how to leverage valuable SharePoint content as a first-class data source for data intelligence and analytics by ingesting it into your lakehouse. Using the new Lakeflow Connect SharePoint connector, we will demonstrate incremental, governed pipelines (via Unity Catalog) that ingest both structured data (Excel, CSV, lists) and unstructured content (PDFs, Office files, images). Data lands as streaming tables enriched with SharePoint metadata to power downstream analytics, search, and AI. We will walk through five low-lift, high-impact patterns: driving an AI-powered BI dashboard from Excel and list data, summarizing and analyzing PDF reports, querying PowerPoint decks via natural language in a secure Genie Room, OCR’ing scanned images for text extraction, and consolidating many CSV reports into a single streaming table.
Talk By: Russ Houberg, Sr. Solutions Architect, Databricks ;

Chapters

FAQs

What is the Lakeflow Connect SharePoint connector?

The Lakeflow Connect SharePoint connector is a Databricks tool that enables incremental, governed ingestion of SharePoint content into the lakehouse as streaming tables. It supports both structured data such as Excel files, CSV files, and lists, as well as unstructured content including PDFs, Office documents, and images. Data is enriched with SharePoint metadata and governed through Unity Catalog so that access controls and lineage are consistently applied downstream.

What are the five GenAI patterns shown in this video?

The five patterns are: OCR processing of historical scanned TIFF documents, an AI-powered BI dashboard driven by Excel and SharePoint list data, AI summarization and analysis of PDF reports, natural language querying of PowerPoint decks via a secure Genie Room, and consolidation of many CSV reports into a single streaming table. Each pattern represents a low-effort, high-impact way to activate SharePoint content for analytics and AI.

How does Unity Catalog govern SharePoint ingestion pipelines?

Unity Catalog provides centralized governance over Lakeflow Connect pipelines, ensuring access controls, data lineage, and metadata are consistently applied as SharePoint data moves into the lakehouse. Permissions established in SharePoint can be reflected downstream in analytics and AI workloads, and governed streaming tables make ingested content available for search, dashboards, and AI use cases. This prevents ungoverned data sprawl as more SharePoint sources are onboarded.

Can Databricks process unstructured SharePoint content like scanned documents?

Yes, this video demonstrates an OCR pipeline that processes historical scanned TIFF documents from SharePoint, extracting text using AI models to make legacy content searchable and analyzable. The pipeline also handles modern PDF and PowerPoint files using Databricks AI functions for parsing and natural language querying. Both structured and unstructured SharePoint content can be ingested and processed within a single Databricks workflow.

Full transcript

[00:10] Okay, welcome in everyone. Thanks for showing up here today um to SharePoint to Lakehouse with LakeFlow Connect top five powerful Genai patterns. I'm Russ Hoberg. Uh I'll be your presenter today. I'm a solutions
[00:22] architect at uh data bricks. been here for three years and uh but but what's interesting is that my background kind of comes into play on this broader story. So um I've been a Microsoft MCP
[00:34] for over 25 years. I was actually a a a SharePoint master back in the 0709 time frame. It's not really a thing anymore. I don't worry about it. Um but but I love SharePoint. I've always enjoyed
[00:47] that technology. And even beyond that, um, my last couple of places of employment also did a lot of enterprise content management and document imaging. So I was really excited to sort of bring
[00:59] together all of my historical capabilities into this one session today. So a couple of housekeeping things. First of all, I appreciate you all showing up. I'll try to keep this as interesting as possible. I know that the uh, you know, lunch is kicking in and
[01:13] the you know, the trip to fans going, everybody might want to take a nap. I get it. Um, I have a lot of demo work to show you today. Uh, there will be a little bit of fumbling around. Um, I couldn't possibly do this without cut
[01:26] and pasting some configuration items. And that's all down there right above the glowy countdown timer to the end of the world. So, um, I just want to make sure that, uh, if if I get hung up, just give me a second and I'll I'll uh I'll
[01:38] try to try to uh get back on track. So, um, all right. Well, let's let's get to it. Okay. So, uh, the agenda today, first of all, I'm going to buzz through, uh, the the slides to a degree. I'm not going to
[01:50] spend a ton of time on them, but the pattern that we're going to be showing you is, uh, essentially a SharePoint library. Um, and the issue is we have a bunch of data in SharePoint, a bunch of different document types. And we need to
[02:04] basically operate over that in an intelligent way. Use that unstructured data as a source and be able to read that in and process that through that uh you know the data bricks pipeline structure and get it into an intelligent
[02:16] portfolio 360 demo or uh uh dashboard. Okay. So pattern one, we're going to have historical data. Think about really old 10ks that were probably scanned in TIFF format a long time ago. Some scans
[02:30] are better than others. They don't always work right. Um, pattern two will be the more modern version of that. It'll be a multimodal set of files. So, there'll be PDFs with charts and graphs. There'll be um maybe investor research
[02:43] uh decks, PowerPoint presentations, so on and so forth. Pattern three will be a lot like autoloader for SharePoint. Essentially, it uses CSV files that uh get appended in uh into SharePoint. Uh
[02:57] we can pull those in and append those into the end of a table. and we'll show you that live. We'll we'll run those processes through and have them output uh changes to our primary dashboard.
[03:09] Pattern four is kind of the money shot. It's the one that everybody asks about. Um uh data bricks now has the ability with LakeFlow Connect, SharePoint connector to communicate directly with Excel spreadsheets inside a document
[03:23] library. And more specifically, I can get all the way down to um an individual worksheet and a cell range. So, we're going to show you that right up front how we can pull that data in and then live later on I will uh I'll actually
[03:36] make some I'll I'll make some changes to a a buy hold recommendation type thing and we'll see that update on the dashboard as well. So, think about your analyst models that uh get these people don't want to get out of the the Excel
[03:48] files and you can still work with that data. And finally, I will circle all the way back up to the um to the dashboard itself, and we'll talk a little bit about it's not a genie demo, but it is always cool when you can talk to your data in the end and ask it questions
[04:01] over things that was that were originally pictures that we scanned in from an archive. So, that's where we're headed today and we're going to be buzzing through here. I will give you just a little bit of the solution story. Again, five different document types
[04:13] piling up across multiple portfolios. We're going to ingest them all through LakeFlow Connect through the bronze, silver, gold medallion architecture and we're going to ingest them into Unity Catalog uh and and and have that catalog
[04:26] awareness. Ultimately that's going to drive a single AIB dashboard a genie space and uh a little sidecar project will be a knowledge assistant that can also be used to feed Genie as well.
[04:40] So just a quick shot of kind of the architecture here. We've got SharePoint with uh multiple folders. We'll we'll show you that in just a moment. It'll have uh the 10Ks, the 10 Q's, and you know, all these earnings reports and so
[04:53] on and so forth. Uh the daily feed is going to bring in things like holdings, pricing uh information, transactions, so that we've got kind of a positions uh model of of our data. And then
[05:05] ultimately that last library that's got those um those analyst Excel models uh and that'll all come in through SharePoint Lakeflow Connect. We'll be ingesting those. Uh we'll be talking
[05:17] about the processing we're going to do there. Uh what's going to be a lot of AI query, AI extraction, AI parse. Uh we'll be pulling some chunks out for the knowledge assistant. We'll be OCRing
[05:29] some of the data. There's just like a ton going on. We're going to show all of that that will ultimately land in bronze and get normalized into silver. So the way that rest of that process looks is once it comes in again like I said we'll
[05:41] be doing AI parse and AI query on our old uh style documents. It's going to be an OCR pipeline that will then be able to read in in uh we'll be able to turn that unstructured data into structured data. Then the filings and IR
[05:55] intelligence pipeline will we'll be doing some chunking and and sentiment analysis there. The daily pipeline feed again it's going to be like autoloadader uh and you're going to see that data append in and it's going to change our portfolio valuation and then the analyst
[06:09] model is is going to come in like we talked about ultimately um what's going to happen here as I go through some of these things is uh we we'll be doing various micro demos uh as as kind of the pipelines are processing and different
[06:21] things like that. Okay. So uh I'm going to tell you about I'm going to show you the SharePoint library real quick. Uh Unity catalog already exists. I am going to show you how to set up the connection to SharePoint as well as then run initial
[06:35] ingestion on on some of the library data again. So I'll just briefly kind of uh go through this, but this is this is what the the historical archive is going to look like. We've got files going back to 2000. We'll be pulling those in.
[06:49] They're going to go through bronze OCR uh AI document parse and ultimately land in our archive financials table. And the purpose of that is to give us that long horizon trend tile inside the dashboard.
[07:03] But to that we will append the uh the some additional data after that. So that again the the land scan uh the data lands in bron. Why does it matter? Um so
[07:15] we're going to see really two ingestion pipelines here. We're going to see one for the bulk of the the the three data types, the first three, and then we'll have a separate ingestion pipeline for Excel. And the and of course the the
[07:29] intention here is to union the old data and the do new data into a single graph. And so here is that new data. It's going to be the earnings intelligence architecture. Uh SharePoint, PDFs, and
[07:41] PPTX again multimodal. We'll use large language models to be able to read and extract that data. uh run it through uh LakeFlow Connect uh historical raw um and then um you know we'll be doing our
[07:53] AI pars. I will dive a little bit into the code. I don't like just scrolling through notebooks. I think that's kind of ugly and boring. So I but I will dive into to two separate notebooks and call
[08:05] out a few different things. Uh I also will show creation of that knowledge agent uh that knowledge assistant in real time. Why does this matter? Um, you know, the
[08:17] important thing here is that we're going to be able to reuse that pipeline for all three use cases. We're actually just refining by the path, the folder, subfolder, as we bring that content in. And we're the real key of this one is
[08:30] really showing off those AI methods as we uh as we get that source data. And we'll also be able to do a SharePoint URL site citation as well as part of the knowledge agent. that daily feed. Here's the the kind of
[08:43] the the last major flow uh other than you know for the the the standard ingestion concepts. We're going to have SharePoint daily uh CSVs that's going to flow through LakeFlow Connect and we will effectively be doing an autoloader
[08:56] like process. It'll have a pointer and it won't do it won't pull in duplicate files. It's only going to pull in the latest. That'll go through our our Spark pipeline uh in order to update our daily positions which will end up driving our
[09:08] um our our portfolio overview tile. We'll we'll show that in a moment. Daily drops to SharePoint. So, I'm actually going to I actually have a little assistant on the side that I'll be running uh you know, I'll let AI
[09:20] progress some things in the background and create me some files and I'll drop those in and that'll be cool and that'll so show off the uh the daily drops into SharePoint and ultimately how that makes it to the live tile.
[09:32] And finally, the last and most probably the coolest demo is going to be, you know, showing how the analyst models can be updated uh by your, you know, by your individual contributors. And when those
[09:44] when those files are changed in SharePoint, we have the ability to detect that change and pull in those differences and uh process that updated data through our our pipeline and ultimately get that into our evaluation
[09:56] snapshot. And that's where I'll be, you know, showing the change from whatever buy to hold and so on and so forth. Okay, I think that takes us to where we're going to stop with the deck. Uh I
[10:10] I wanted to buzz through that. I'm not I'm not all about living in PowerPoint. So, um but yes, I'm going to show uh the the I'm going to be going back and forth to this uh to this PowerPoint or to this uh uh dashboard over and over again, and
[10:23] we'll see the changes happening. Got a little far there. All right. So, let's change. Come on. There it goes.
[10:37] Okay. So, first SharePoint. Let's take a look at what we have here. So, here we have a SharePoint library. It has a historical archive. We talked about that. If we go in and take a look at it, um we're going to see our different uh you know, folks
[10:50] that we have as part of our portfolio. We're going to have these various files that you know, you might be able to tell that this is kind of an old scan and and it's uh you know, it's it's kind of it's a little grainy. It's not too bad. We
[11:02] shouldn't have too much trouble with that one. But you'll see uh those scan TIFF files come from that library from that subfolder. We'll also have earnings intelligence reports uh 10Ks and 10 Q's and again these are going to be our
[11:15] PowerPoint presentations and our PDFs again multimodal. And then we'll also have our daily feeds. We're which are going to be our CSVs for holdings, pricing, and trades. And as that data updates, we're going to
[11:27] see that affect our portfolio. And then we have our our analyst model. So you'll see me come in and and modify these files just to to level set. I'm going to open up Nvidia today. Nvidia says it's a hold at 682. We'll keep that in mind. Um
[11:42] and so that's where all of the data is coming from. So let's take a quick gander at the dashboard as it is today. Okay, I hope everybody can see most of that uh fairly well. Um I'm I'm a little limited on on the quality and the and
[11:55] the the size of everything. So I'll I'll I'll try to zoom in where I can. Um but today what what I want you to notice from here is we're starting with 479 million uh in total market value and 6.54 million in late uh latest daily
[12:09] P&L. Um for the record, this is not trading advice. This is all simulated generated data. None of this is real. Uh so just pay no attention to the man behind the curtain there. And then and
[12:21] then the portfolio overview will update. The total market value will update and the latest P&L will update as our dailies come in. And then our um the other thing we're going to see here is currently Nvidia is that hold at 682. So
[12:35] that's where we saw that in the spreadsheet. That's where we currently have everything starting from. So let's go ahead and and dig into this thing and and start uh start showing off. So, first thing I want to show you, and again, it's a little bit small, but
[12:48] let's uh let's see what we can do here. So, first thing I want to do is go into our catalog, and I want to create a SharePoint connection. So, I'm going to go create connection and connection name is
[13:02] uh live sp and then the connection type is going to be SharePoint. So, we'll click that there. and the off type is going to be user to machine. And
[13:15] we'll click next. Now, we need a couple of bits of information here. The first bit of information that you need is the URL to your uh your loca your you know your your site collection. You also need
[13:27] your tenant ID. Now, the tenant ID is something you can go to Azure and get into and get from Entra um the actual tenant ID. So, we'll we'll pop that in here. And this is fortunately just a tear down tenant. So, it's not going to
[13:40] matter. or you guys aren't going to be able to do anything with it. Uh uh so we'll go ahead and we'll do a create connection. And so the off succeeded, it's in it already worked. We already have a connection and and and creating that established connection is really
[13:53] simple. Now the cool thing here is of course that the connection itself is governed by Unity catalog. So if that connection is able to get some to some sensitive data, we can control what service principle or individual you know roles and groups can get to and even use
[14:07] that connection. So, moving on. Let me go into our ability to let's get here and go in. I'm going to do a new Now, I'm going to go bring some
[14:20] data in. So, let's go uh we'll do add or upload data. And I'm going to just do this with Excel just so that we can get right to the cool stuff. So, we'll start in a really cool place. So, live
[14:33] um ingestion pipe. I'm just going to put this in a a test catalog and this Excel test schema and I'm put pulling in. So we have two different types of files we can pull in
[14:45] here. The one is going to be structured tables. So that's what we're doing to here with Excel, but the other things that have structure are going to be like CSVs and JSON files uh and so on and so forth, right? Um the unstructured data would be if we want to pull in pictures.
[14:58] Maybe you're in doing insurance and you want to pull in pictures of uh crashed vehicle or something and while you're doing claims processing. Uh so unstructured data could be uh you know PDFs and and images and docs and so on
[15:10] and so forth. So we'll pull in some structured data. We'll create the pipeline. Now I'm going to do something a little bit interesting here and I'm going to give you a little bit of unvarnished truth. Um at this point we want to give
[15:22] this a folder where the library is that we want to pull the data from. And in and actually that's not the folder I want. I want this one. And the reason why we want this to be a subfolder is
[15:36] okay hang on. So get that copied. The reason we want to bring in a specific subfolder is that we have to make um some graph API calls to go get the actual documents that are in
[15:49] there. And if we set it at just the root SharePoint library, it's going to take quite a while to go literally crawl that whole tree structure and then and then give you the option as as to what you want to bring in. So, in this case, I
[16:01] just gave it the subfolder for those Excel analyst files, and it should come in reasonably quickly. There it is right there. And now, oh, and see, keeps flipping back and forth, but that's okay. So, normally you you're going to have the opportunity to um basically
[16:16] pick individual files here if you wanted to. If we select the folder, it will process all the files. So, I'll just go ahead and do that here. I'll select the file format as Excel. And then it'll ask me, well, what do you want to do with
[16:29] that data? What I want to bring in is the cell range A8 through B14. And the sheet name is actually DCF.
[16:41] And we'll see this a little bit more later. So, what that's going to do is it's going to process all of those um Excel files and it's going to pull that exact block from that exact worksheet on all of those Excel uh documents. So, the
[16:55] the next thing it's doing, it's initializing the pipeline, kind of getting some things wired up in the background. It's going to bring up a scheduling agent for how often to run. You can run these as frequently or as infrequently as you like. It could be by
[17:07] the minute, five minutes, 15, half hour, once a day, whatever you want to do. I'm actually going to kill the schedule when it pops up because it just burns compute that I don't need because this is just a one-time demo. So, I'm just going to delete the schedule. I only need this to
[17:19] run once. So, I will save and run the pipeline. And this is actually going to go faster than you would think that it would. Um, famous last words, right? I mean, now that I've said that, it's probably going
[17:31] to take forever. Uh, but anyway, so, uh, this is going to Okay, there. There it goes. It's already getting the tables ready, and we've got the streaming tables. At a few moments here, we'll see some output records start to pop. And that is the actual processing of the
[17:44] individual um the individual cell cell chunks. So, that's done. And this is almost finished. It's running. It's completed. And we're finished. It was that fast. So, we're going to view in the catalog this data. And I want to
[17:56] show you a couple of things here. So inside this uh library as soon or in this uh uh uh table is the actual ingested location for each of these
[18:08] different documents. So the first thing I want to show you is the SharePoint metadata. That's really hard to see. Uh let's see if we can kind of grow that a bit. So one thing I wanted to call out is if you have custom metadata inside
[18:20] your SharePoint library, for example, in document imaging circles, a lot of times you'll see a document type. This is where we're going to call it a 10K or an analyst document or a 10Q. You'll get that dock type value inside this additional metadata field. Of course,
[18:32] you get all the other, you know, details, who created it, the size, and all those things. Later on, we're going to use the actual uh file name as as part of our citation source in one of the other demos. And then what you're
[18:46] going to see is the uh the path to the document and and the name. And that's and that's really kind of it. the the rest of that data is going to be, you know, it's going to be part of the the
[18:58] the binary content that that kind of came in. And so I kind of pop this out here, you kind of start to see where all that data came in. But what matters is that this table actually has those values. So if I take a look at
[19:12] that actual document, if I can find it. So in that range for Nvidia for example on the DCF we'll we're going to be look we're going to see terminal um we're
[19:25] going to see you know a couple of other things here uh you know rating anchor all these these are going to be the the the fields it's sort of a key value pair and this is the value so then when I get in on this side and actually look at the
[19:37] document that that came in. So there's those terminals and the fair values and all that stuff and then you get the actual value that came in. So all that just came from Excel. So that's just the like how you get it done, but uh we're going to take this a step further.
[19:51] So what I'd like to do now is just kind of show off the jobs and pipelines that are kind of behind this demo. There's that live ingestion pipe we just created. And all of these pipelines work together to process that data in from
[20:03] SharePoint, go through our uh you know, bronze, silver, gold pattern, and do the OCR and all these wonderful things. So let me just show you what these things do. So first one is our historical pipe. This brings the data in from uh
[20:15] SharePoint and this is going to be those unstructured files into SharePoint historical raw. That's all it does. It's just our LakeFlow uh uh connect lakeflow connect pipeline to SharePoint. The next one that's kind of interesting
[20:27] is going to be the OCR pipe. And this is the first one I'm going to dive in on. So once you have that data in, we've got this these TIFF files stored in binary in in a in a uh in a column. And so we
[20:39] can pass that along and now we can do some interesting things with it. So again, I'm I'm going to sort of scroll down to the part of the code that matters. And I'm not going to just buzz through this whole thing. Um, by the
[20:51] way, I intended to create a a QR code at the end for the um for like my repo and the and the and the and this deck. I I I failed to do that. I apologize. But I will be able all this code is is we can
[21:03] get this out if you guys want it. So, what I wanted to show you though was a couple of things. I don't know why I have this artifact on here, but that's okay. So, a couple things I wanted to show you. So, first of all, we've got our um
[21:17] essentially our our our different prompts and what you're what they're going to be doing. There's uh the various different components. And then you'll see uh what we're doing is we're we're pulling in the uh the historical
[21:29] data. We're telling it to to essentially OCR the data. And we want it to output in this very specific pattern that we've provided. That data is going to come in uh you know and we're going to get all the the data values out exactly where
[21:42] where we ask and and it's going to it's going to come in through the OCR. So that's the first one. The second one will be part of the AI extraction pipe. And this is going to be a little bit more interesting. So the data that comes
[21:55] in, we're going to actually do this a little bit more. There's going to be a lot more to this one. So let's jump into the code real quick. And I just want to show you again, we're going to be doing some prompts. Um, you know, here's a chunk uh prompt and then there's going
[22:08] to be uh let's see here. Let's keep going. Let's get down to the actual All right. So, here's the first here's one of the the the prompts. Get us exactly out where we want. Where is my
[22:29] This is too big. All right. Where's the AI parse?
[22:41] All right. So, AI parse document. Here we go. So, AI parse is going to chain into AI query and in some cases it'll chain into sentiment analysis uh analyze sentiment and in other cases it'll it'll
[22:53] chain into AI extract which is going to pull the chunks out that allow me to do the knowledge assistant. So essentially you're starting with raw binary and we're passing that in to datab bricks essentially AI document im in uh
[23:06] document ingestion and intelligence functions to get that rich data out and it's and it's actually quite accurate. I'm actually going to demo that in a moment and show you what that looks like graphically as we go through it. So then
[23:18] we have daily feed pipe. This is again going to pull in I'm not going to dive into any more notebooks but this is going to pull in all those three different files as they get generated. There's the Excel ingest pipe. We're going to see this fire uh again as we as
[23:30] we show that in real time. And finally, we've got our our gold uh silver to gold pipe, which ultimately ends up populating our dashboard. Okay, so let's do the first one. Um the first part of
[23:42] the demo is going to be the um it's going to be the the uh the daily feed pipe. So, I'm going to go over to my lovely AI assistant over here, and I'm going to fire on my 01 progress
[23:55] daily skill. And we'll give that I'm gonna I'm gonna I'm going to go back and forth. We'll make sure that the the AI is happy. And ultimately what this is going to do for
[24:07] me is it's going to give me some new files first. So my my most recent day in SharePoint. So if I go look at SharePoint real
[24:19] quick. So if I Yeah. Yeah. No, I'm getting there. So, we're we're in SharePoint. Yeah, you won't be able to see what I got going on the side there. No worries. Uh that this is all again the lovely
[24:32] behind the scenes assistant. You know, you don't want to see that. So, historical archive. Uh actually, that's not what we want. We want the daily feeds. So, if we look at holdings, we'll see our most recent day here is actually
[24:44] going to be 529. So, this will flip it over to 530. And
[24:57] if I can get my feed to go. Oh no. Good times. I have to reoff my token.
[25:10] Yes, this is going to be great. I just did this this morning and it and uh so it must have just timed out. All right, so let me get let me get my off going. Slight detour here.
[25:34] There we go. Here comes myert. You guys don't need to see all that, but you are. It's fine. Not going to hurt anything. Just going to log in and then we'll be able to do stuff. It's going to be cool.
[25:47] All right. Token authenticated. Go. All right. Again, so what we're waiting on is for the AI to build my my daily my daily churn. Essentially, it's going to create me um three new files that should
[25:59] be five should end in 5:30 and then I will end up dragging those into these three folders uh as Oh, we already got it. It's going to start with 601. So, it went to the next business day. So, here's 601 and that is for trades. We'll
[26:13] drop that in there. And then prices 601 comes in here. And then holding 601 goes in there.
[26:25] Okay. So we will now tell the AI to proceed and it's going to fire off my uh you know my pipelines in order. I could have just left them run but I have a habit of just leaving them run and then they don't like it cuz you're burning
[26:37] cash and you don't really need that. So uh we're going to just let the AI uh fire off the pipes. So, we'll go take a look at our pipelines. And pretty soon here, we'll see this SharePoint historical pipe fire. And when I see
[26:51] that, then I'll know it's off and running. Um, and we will be good to go. So, keep the AI rolling. And then what that does for me is it gives me the opportunity to run a secondary little
[27:03] side quest demo while those pipes get get moving. I can keep going. Got to have the approvals. Some things you just can't auto approve. All right. So, the pipe should be going soon.
[27:16] Firing the chain. So, we should start seeing this move. And once I see that move, I'll be I'll be happy to uh to move on. This is the fun with uh pipeline-based demos. Got to get the pipelines to fire
[27:30] in order. How's everybody doing today? Good. Everybody looking forward to the chain smokers tonight? It's going to be good. It's awesome.
[27:53] All right, pipe's going. So, I'm going to I'm going to get going here real quick. I'm starting to run out of time. So, um while this while this is firing in the background, what I'm going to do real quick is I'm going to go over to my data bricks location. I'm going to go to
[28:06] agents and I'm just going to kind of show you what it looks like for the uh to to to have document parsing. I told you I was going to show you that. So over here I have an investor day deck
[28:19] which I will drop in and well we're going to see this work. So this what this is is essentially document ingestion. So this is exactly what's happening under the covers in those methods. So, what's happening is
[28:31] the the document's going to come in and it's going to render first and then it's going to start to process all of those individual documents. You will start to see some bounding boxes automatically appear and it's going to extract all that data out and put it into a JSON
[28:44] format. So, there's the bounding boxes and it's all coming down and it's just processing the slides in order and as it kind of comes down we're going to get all this rich information. Keep going.
[28:56] Page six, page seven. We're going to keep parsing until we get to the end. There's 10 pages and we'll we'll start seeing more bounding boxes
[29:23] So, let's go back to here. And I think we have most of the information. We'll just go ahead and take a look at it uh the way it is. But essentially, you know, you get all that data process in. And now we have the
[29:36] ability to go do something with that, right? We can we can send that through our pipeline and get all that rich content. So that's kind of how that works. So we're going to go back to our jobs and pipelines. That one's firing. It should be just about over. I'll go take
[29:49] a look real quick. See how we're doing. Okay, that's just about done. We'll start seeing some output rows come in from those three new files.
[30:08] So there we we're getting our output there.
[30:20] Almost done. So moving on. Come back here. So that is just about finished.
[30:32] Okay, it succeeded. Got it. So now we should be moving on to our gold layer which is going to update our actual dashboard. And this doesn't take long to run. So we'll get into the gold dashboard pipeline.
[30:44] Take a look at our DAG here in a moment. This works a lot faster at home.
[30:57] Okay. So, I'm going to zoom in on this just a little bit. So, we're we're processing essentially the the kind of the the key concept of this demo is that we're bringing in multiple unstructured sources. We are then processing those through through you know uh uh Spark
[31:10] declarative pipelines and then we end up joining those in to materialize views that ultimately drive our fundamentals, our sentiment, our valuation snapshot and ultimately our gold sector rollup.
[31:22] So we're done with the first layer, we're into the second and now we're all going to start feeding everything into our gold sector rollup and then this will be done.
[31:41] Okay, I think we're just about done. Okay, succeeded. So, let's go to the dashboard. So, now what we're going to see is 47909 and 6.54. If the demo gods are with me, they will update to new values.
[32:03] >> Hey, we're losing money, but at least it updated. Awesome. So, and then now I'm going to do one more quick change here. Let's go take a look. Um, again, uh, we'll we'll we'll go start off by taking a a gander at
[32:17] Actually, I'm going to fire this guy off. We're going to go O2 and get him rolling since this is taking a while. We're going to move our our daily forward. So, we're going to flip that rating. Okay.
[32:29] So for our analyst model that's going to crunch in the background while I take a look at this. All right. So analyst, we're going to be playing around with Nvidia today. Our our uh Nvidia uh analyst is is come in
[32:43] and he's bas and we're basically saying that um you know we've got this hold uh at 682. So we're going to go ahead and generate a new document,
[32:57] an updated version of Nvidia, and it is now done. So, I'm going to close this and I'm going to bring in my updated Nvidia document. I'm going to drop it in here and I'm going to do a replace. And so, that's going to again fire off. It's
[33:09] essentially going to be a trigger so that the next time that my pipeline runs, which hopefully can happen faster than last time, we'll go ahead and get that over here. Get it looking.
[33:23] This will be probably the towards the end here. And then while that's running, I do have one other side demo which we should have enough time for. Um, probably not going to have a ton of time for questions afterwards, but what we can do is there's two more
[33:35] opportunities. So after the session's over, there'll be 15 minutes or so where we can have conversations. I also have one slot left in a brain date for tomorrow specifically for this session. And if if that fills up, I can do a
[33:48] second brain date as well and and fire up another one if folks have questions. So based I'll I'll I'll gauge that based on how much you know how many questions there are after we leave here because the countdown typer to Doom is telling
[34:00] me I have six minutes left. So okay so this is going. We're going to shortly Okay. So we've already got our our pipelines are running. While that's happening I'm going to go back into the workspace. I'm going to go back to my
[34:15] agents and I'm going to go create a knowledge assistant. Get that queued up down here.
[34:29] Daily feeds analyst knowledge assistant. Here's the detail. Okay. So, let's go create a knowledge assistant. We're going to create a knowledge assistant based on a vector search index that was created as part of those pipelines. That
[34:41] was actually a workflow. It's one of the things I didn't show you, but we did do that. Um, so then we're going to go go ahead and paste in a
[34:53] description. I'm going to get this from a search index. It's going to be in our silver uh area or silver schema and it's going to be our filings index and it's going to pull from. So the the the
[35:06] citation is going to come from the file name, but the text is going to come from the chunk text. And the last thing I'm going to do here is I am going to give it some instructions and basically tell it how to function.
[35:20] So we'll copy that. We'll bring it over here. And we're going to say you are an analyst assistant um answering questions about a portfolio and some publicly listed issuers, 10 Q's, 10Ks, so on and
[35:33] so forth. But the cool thing is so I'm sure you all have seen the Genie demos and and I'll if we have time at the end I'll I'll show you one as well. But that that has to go through and look at context and think about things. With the knowledge assistant, we already have the
[35:47] the vector index already set up. So I can go in and just ask a question and it comes pretty quickly. It doesn't take long to get the answer back. So what have issuers said about AI capital expenditure pressure in their most
[36:00] recent investor day decks? Summarize particular and site the source. So that investor day deck that came in the PowerPoint files that is all now been chunked and it is and it is put into our vector uh vector store uh vector index
[36:14] and so that gives us the ability to ask this question and get a get an answer really really quickly. So let's go take a look at our pipeline real quick. Um gold is still running. Let's go take a look real quick and see where we're at.
[36:26] Okay, so we are on our last little bit. This is great. So this is going to land just right. So, we're done there. Now, let's take a look before we go into the
[36:39] actual dashboard. So, what did it change to? So, now we're at buy for 805. So, if I come into the dashboard and I come down to my valuation snapshot, we should see Nvidia pop to the buy side up top
[36:53] here and go to buy at at 805, I believe. So, let's refresh this and we'll watch Nvidia change now that all that's been processed. There's Nvidia. It's a buy. It's 805. And that was all based on the
[37:07] analyst inputting into an Excel spreadsheet. And we went and said, "Okay, fine. Keep your spreadsheet. We'll we'll go ahead and pull the data in anyway." So, that's really kind of the key uh component of the demo. We
[37:20] have two minutes left. The last thing I'll just show you is that of course we have a genie room. Of course, we have the ability to ask natural language questions of our data. And we can come in and do Oh, let me let me grab my my
[37:35] prettiest output question. And we'll drop that in here. And we'll hit paste. And we'll say uh I want to know what the market value distribution of GI uh by by GICS sector
[37:49] and ticker in the portfolio. So, I didn't have it broken down by sector and and I wanted some additional information. And so again, all this unstructured data comes in. Um, my my friends at my my customer, many of which
[38:01] are here today, will tell you that it takes a while to get the a dashboard approved and updated and through the system. But if you can just come in and ask the dashboard questions, you don't have this problem of waiting on someone to update the dashboard. So, this will
[38:15] go through and it'll think a little bit. We can look at its chain of thought and it's it's grabbing some data and it's it's done a natural language transition. And here's the code if you want to see it, but we have the ability now to get
[38:27] that answer by by sector and uh and and and it'll even color code it and and show us where all of our our our positions are. So, uh that's pretty much it. That's all I have. I've got a minute and a half. If anybody has any burning
[38:40] questions, I'm happy to to address. Otherwise, I'll I'll probably just do a wrap.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.