Team DataFam Tableau User Group 16 Jan 2026
"How to Leverage AI for conversational Dashboard using MCP" by Harim Jung
40 min 6,659 words transcribed
Summary
Team DataFam TUG meeting 55, hosted by Chris McClellan, features Harim Jung, Tableau Public Ambassador and founder of the Visual Climate, on leveraging AI for conversational dashboards with LangChain and MCP. Building on Will Sutton's GitHub resources and videos, Harim walks through how a LangChain agent orchestrates between a user's question and Tableau Cloud: the agent picks a tool, reads the data source's metadata through the Metadata (GraphQL) API using its LUID, maps the question to fields in a data dictionary, then queries the data through the VizQL Data Service. LangSmith traces each step as a waterfall, with latency and cost per question. She shows the setup (an environment file holding the LLM key, LangSmith key, Tableau Cloud credentials and data source LUID), the agent answering a question about China's CO2 emissions in a terminal and in a web chat app, and notes the model can be OpenAI, Gemini, Llama or DeepSeek. Asking 'so what?', she then presents a RAG platform for climate data: unstructured PDFs parsed with LlamaParse and PyPDFLoader, local embeddings stored in Chroma DB, Gemini for reasoning, and Tableau as the end product, with strategic briefings and micro-citations to guard against hallucination. Her case for Tableau in the AI era is that AI tells you the what, while visualisation delivers the so what. In Q&A, Wendy Small asks whether proprietary data could leak to the LLM provider and notes that a personal access token carries its user's permissions to everyone using the bot, and Ilija Stojić asks how to stop an end user querying a data source they couldn't open in Tableau. Harim also answers whether this works with Tableau Server as well as Cloud. The meeting closes with the audience survey, ideas for future topics and presenters, and a preview of Tore Levinsen and Steve Adams on dashboard templating.
Listen to this TUG
Prefer to listen? This meeting is also an episode of the Tableau TUG recordings podcast.
Key points
- A LangChain agent sits between the user and Tableau Cloud: it picks a tool, reads the data source's metadata, then queries the data.
- The Metadata (GraphQL) API supplies field definitions keyed on the data source LUID; the VizQL Data Service returns the actual data.
- LangSmith traces every step as a waterfall, so you can see latency and cost for each question.
- The LLM is swappable: OpenAI, Gemini, Llama or DeepSeek can sit behind the same agent.
- A RAG pipeline can turn unstructured PDFs into structured data whose end product is a Tableau dashboard, with citations to check the AI's answers.
- Harim's case for Tableau in the AI era: AI provides the what, visualisation provides the so what.
- A personal access token gives everyone using the chatbot the permissions of the user who created it.
- Restricting which data sources an end user can query through a chatbot is still an open question worth designing for.
Jump to a chapter (21)
- 0:00 Welcome to Team DataFam TUG 55
- 0:46 Harim Jung: AI for conversational dashboards
- 1:46 When Tableau meets AI
- 3:12 How LangChain and MCP connect to Tableau
- 4:09 Tracing and cost monitoring with LangSmith
- 5:14 How the agent chooses a tool
- 7:13 From question to answer
- 10:19 Connecting Tableau Cloud to an LLM
- 11:43 Demo: the agent and the web app
- 13:27 So what? Beyond the conversational dashboard
- 14:05 A Tableau-driven RAG platform
- 15:24 A privacy-first data pipeline
- 17:06 Why we still need Tableau
- 18:41 A RAG-driven climate dashboard
- 21:17 Q&A: is proprietary data safe?
- 25:54 Q&A: restricting what end users can query
- 28:29 Q&A: slides, and Tableau Server or Cloud
- 28:56 Audience survey
- 34:46 Future topics and presenters
- 38:36 Next time: Tore and Steve on templating
- 39:30 Thanks and close
Transcript
0:00 Chris McClellan: Welcome to Team DataFam Tableau User Group meeting number fifty-five. This TUG works a little bit differently to others because we aim to capture all the in-person and unrecorded presentations and give them out to a wider audience.
0:14 Uh, I don't really care if you've presented at a big conference like TC, or some smaller meetings, or if you've never presented before. If you've got some Tableau-related content, please get in touch. We can find a spot for you. My email is team.DataFam@Gmail.com.
0:31 But today we've got Harim talking about how to leverage AI for conversational dashboards using MCP.
0:38 Something about Tableau, something about Streamlit.
0:41 I don't know. Everyone's here to learn. So I'll hand it over to Harim, and she can take it from here.
0:47 Harim Jung: All right. Okay. Let's start it. So once again, hello, everyone. Thank you for having me here today. This is Harim. I'm a Tableau Public Ambassador and a founder of the Visual Climate. So today, I want to show you how we are actually redefining data interaction by leveraging the AI for the conversational dashboard using the LangChain-MCP. I believe you already heard about the LangChain-MCP, MCP so many times from here and there, and I want to give you, like, how you can actually leverage using this reference as an idea.
1:17 So we are moving beyond a simple chart to a Tableau-driven RAG platform in the end. So let me get started.
1:25 So Tableau is already, uh, powerful, but we must ask ourself, "What happens when Tableau meets AI?" This is a major question I've actually got requested so many times because lots of the data people or non-data people, they feel like Tableau is nonetheless and not like it before, because so many AI things are already supporting us in a good way. But we are not just in adding a chatbot in Tableau, right?
1:50 Tableau and dashboard is not a simple dashboard when it is only showing the visual impact. Actually, what we need to give them is, like, a data to make it more accessible and more actionable. That is the goal of, like, Tableau from the start.
2:05 So our goal is to build the first AI-driven climate dashboard library. I've been talking about this one, um, almost, like, for two years, and I've been developing with so many, like, a climate data engine using the Tableau AI and RAG system. And while the human intuition and visual exploration are vital, but we need to speed and beauty of, like, Tableau combined with AI, like, in real-time monitoring as well.
2:31 So you can actually imagine if Tableau could answer you back. That is really powerful. That is actually our dream. That's why Tableau give us, like, agentic experience and some kind of, like, generative experience. But still, I believe you might feel a little bit thirsty because there's a kind of, like, a great opportunity, but at the same time, there's a limit of the challenge.
2:51 So today, let's check how we actually enable natural language query and automated data interpretation and conversational analytic as well. So this allow any user ask complex question, and you can actually apply this idea and concept to your Tableau and LLM and AI system as well. So let's start it.
3:12 So to understand how this actually work, you need to know how the LangChain or how the LLM system is actually working like a banker.
3:22 It acts as orchestrator between the user internet and, like, a data location as well. So let's go very simple. I'm a big fan of, like, Tableau, like, visionary Will Sutton, and also the great, like, Tableau, like, resource, like MCP or something. You can actually follow him, and you can find the GitHub resource, very, very good one. And I'm the one who actually learned this kind of a LangChain and MCP idea from his cour- his courses from the YouTube.
3:48 So it's very simple. Using the user query, you can connect to LLM chain and then connect to the agent. Finally, you… we can connect to the tool for the Tableau VizQL API and then make your own output.
4:01 So finally, we need to understand, like, what is the VizQL API tool is working for your dashboard, and I will go through really a glimpse way.
4:10 So, uh, I will not go through very in detail how to do that because the Git repository, like, it included everything in detail, and you can get a very, very detailed information. But what I want to tell you is I use LangSmith for the tracing and cost monitoring from the start because LangChain is providing a very good interface, and we are data people, we are visual people, so whenever we see the dashboard, we can intuitively understand, "What am I doing?
4:36 How much am I spending?" So for example, if a user or myself ask a Tableau, a LLM d- d- driven Tableau, like, specific question like, "What is the CO2 emission 2010 in China?" Get all of the information. Then you can see the exact timestamp and latency in here and specific data set and also, like, how much it cost you actually did it, uh, require.
4:58 So we can see the latency and cost. The cost is really actually small, but you can actually being obsessed, right?
5:05 The idea of, like, a LangSmith is really beautiful. And then I will go through, like, how the agent decision process is working inside of a LangChain MCP process. So the system is used like a waterfall trace to decide what to do.
5:20 As, as long as here, you can see the agent determines if the question is simple, then normally they decide if you use, like, a very simple, like, a tool, like, a simple data source QA tool. The good part of, like, using the MCP, using the, like, LangChain system or any kind of, like, API thing is you don't need to know everything about the AI. You know, you…
5:40 What you need to know is, like, how the architecture is working on, and then, like, ask the right question. Then AI system, LangChain or MCP, they figure out what kind of tool, like, based on, like, a latest version, like, a function, and they match it up. So I just, I simply ask her a simple question, and then it simply directly mention the simple data source query tool, and then read the metadata via- GraphQL API, and then execute it for you through the VizQL API.
6:08 So for instance, user ask the natural language, like, "What is the emission of, like, CO2?" And this is very simple question, so LangChain prompt, uh, directly, uh, connect to the LLM chain and then make a prompt using their own agent and just select the tool. And then after that tool, we can connect to a Tableau Cloud, which like GraphQL and VizQL as well.
6:31 The difference between a GraphQL is, like, GraphQL is reading the metadata using, like, LUID and field definition, because we need to actually connect our data sources to LLM in a natural language. And then using the VizQL API is like exact data query. So VizQL API is actually retouching your dashboard, and then if the user asks specific question, then the, like, visually appealing impact is actually, uh, worked by the VizQL API.
6:57 So after that, LangChain return the structured answer as well. So you can see the answer, you can see the, like, answer to require like a dashboard impact as well using this system. So this is very sim… I, I know this is not very simple, but idea is very simple actually.
7:12 So after that, the, this is the stage where we run the app.
7:19 Hi. The human input, like calculate the CO2 emission, like, uh, China in twenty-ten. I just, like, simply ask it, right? And then it's sent to the agent, LangChain agent, and then begin to the process like translating intent or data-driven responses. So you can see the… Right after that, you can see the output in here. I ask a natural language process, so it designed specifically, like, agent VizQL.
7:44 This AI analyst edits this context based on only about the connected LUID. So it connects instruction is like probably types, simple data source QA, I already mentioned about the function tool, and then strict to the data source, like, provided ensuring high accuracy and zero hallucination is really important. So long story short, I know this is a little bit complicated, but you don't need to do anything.
8:06 Well, based on the code, you can ask a specific question, then LangChain prompt and LangChain agent, like, figure out what kind of tool they're gonna use, and then you can see the ChatGPT is already, like, connected, so it make their own prompt by their, like, agent system. So after that, we can mapping out to the data dictionary, because end up we need to read our dashboard connected to our, like, question, right?
8:31 So LangChain, like, automatically mapping to the data dictionary like this. So AI analyzes a prompt by the mapping the data dictionary, and it defines fields like emission in a year and country. Do you remember, do you remember what question I asked? The CO2 emission for the China in twenty-ten. And interestingly, this one already understand what is the question, and then mapping the data source, and then figure out, like, what kind of fields do we have to match and to answer this question.
8:59 So it figure out, like, emission and year and country, and then understand the data type. Like, is it, like, a string data source, like, in data source, is it, like, a real or not? And then ensure that this tool query is synthetically correct.
9:13 So this is a structure of, like, how the LangChain is working, and you can see that I, I connected, like, a ChatGPT model, so which like, uh, OpenAI, like, LLM. So it connected OpenAI LLM, but doesn't have to be ours, like, OpenAI. You can connect to the, like, Gemini or you can connect to Llama or like DeepSeek. Any kind of model, you can actually do that.
9:34 And then finally we can get our output. Very interesting. So the good part of, like, the LangSmith and good part of, like, the dashboard is show like a waterfall. You can see what's actually going on inside of, like, a step-by-step, and you can see the output as a, as a, like, terminal value. But I understand this is quite complicated, so that's why we connect this one in a, like, a reconnect to the, like, a web host to make a real conversational chatbot inside of a dashboard.
10:00 That's, that's why we are learning. We are, like, using the LLM connection and make the dashboard fully conversationer.
10:10 Okay, so this is actually the very glimpse of, like, how the AI structure is working, and then I will go for, like, a quick learning, like, quick learning demo. So we can now look at how we actually connect the Tableau to a LLM.
10:26 I already told you, like, the, all of the sources actually from, we can find it easily from the GitHub resource, from the Tableau or Salesforce or, like, all sudden. But all you need is actually the end file containing your, like, LLM, like, uh, LLM to– LLM, OpenAI LLM token or, like, a LangSmith API key along with your, like, a public credential and specific data source like LUID.
10:50 Because in the unit to connect your data source to the LLM, right? And the key is actually LUID. And then you can run the application like a testing mode via command line and launch the full web interface mode as well. So again, if you figure out, like, there's a kind of, like, complexity inside of, like, this code, it's only like a empty template file.
11:11 Be careful, like, to get your API key and get your API key for any LLM model and try to open your cloud. You know, you can actually get the free cloud. Click match the client ID and secret ID to enable, and then get the LUID for the GraphQL. That's all. This is a very essential things, and you can definitely find this is a little bit hard, but after you get used to, you will have a like an, uh, like another level of opportunity to connect your Tableau data and Tableau and LLM as well.
11:42 So, so this is actually the agent start working with the Tableau Cloud. So you– whenever you run like a main Python code and then reconnect to the, like, app Python code, and then you can connect to Tableau Cloud AI and LangSmith trace all of, like, journey this one. I know this is quite small, and it's really hard to catch it up, but the idea is very simple.
12:05 The video demonstrate like agent working as a live, and we call this API code, which trigger the Tableau Cloud API, and then you can see the LangChain trace happening in the background and the dashboard updated dynamically based on the conversation. So I just say run this code, and you can figure out like what is going on inside of like this terminal, right?
12:26 Human messages here. I already asked for like specific question, and it will give you the answer. Like, um, the user input is this unique. It, uh, interpreted my question as a user input as a different way. And then after this, this code is, like, running, you can connect to the web post as well.
12:46 So I will open this one. This is finally the web post. You can actually use it like a chatbot and connect to the Tableau dashboard or yours one or the other one's. But the good part is you can ask any kind of question based on your Tableau dashboard, based on your Tableau Cloud. I simply ask about the CO2 emission of the China, but interestingly, it give us answer not even of China, Taiwan, and Hong Kong, which I know this is a little bit like sensitive, but this is how I feel, oh, the LLM is actually working because I simply asked China, but it read a, like a better context about like a, like a different, uh, concept of like China as well.
13:22 So you can connect it as well.
13:26 And okay. And after that, you might feel wonder. So okay, I really like this idea. I know how to… I know like people want to make a conversational dashboard, but so what? Actually, that is my question as well because I know I love my conversational dashboard, but that's not enough, you know? No matter how much I like can ask for like dashboard to reflect, there's nothing.
13:48 Like, I cannot say nothing, but it is not like a big impact I can actually get it on. So I thought and I asked myself, "How can I actually leverage AI connecting to Tableau and then make a real product that no one actually thought about it, but we finally think we can actually make it?" So let's go further. So I build a RAG platform, like Retrieval-Augmented Generation, like connected directly with the Tableau.
14:12 The meaning of the RAG is very– the RAG system is like very varied, but in here, my concept is very simple. I want to actually like gather all of kind of unstructured data source, like a PDF document or all related data source, and then connect them into the LlamaParse and structural, structure this unstructured upload data source, and then keep them as a DV and connect this data source using the LLM model, using the…
14:41 I used the Gemini Pro, but it already like, uh, updated like Pro two or three. This is very simple. Using the da- uh, structure the unstructured data store, connect to the LLM, and then these processes exactly the same. Connecting to the data source like LLM and then like a BigQuery, and then connect to the API bridge, and then finally make the dashboards working together with the RAG system.
15:08 So finally, I could make my own RAG system, and then the final product is designed and developed by the Tableau, which is very powerful because we cannot rely on only the texture or like only a whole like table thing. So I want to give you a very glimpse like how this privacy, like privacy, uh, privacy first data pipeline is working. So the for security, I use like a local security parameter, like I ingest a PDF mostly.
15:40 So I use like a PyPDFLoader, a like very famous like library, right? And then like I use local embedding models, and then store them in like a Chroma DB. You don't need to follow Chroma DB or like a, like PyLoader, PyPDFLoader, but I simply love this way because this is easy and manage to like you're embedding your own local system. And then I, I mentioned like a Gemini one point five, like a two or a three, just because I need to reasoning like a engine that will digest the data set and like connect to the LLM.
16:11 And then inside of a RAG system, I don't want to make any hallucination, but I only want to get an answer based on the this RAG system, the data source I already give as like a input. And then this is really interesting. After all of the like RAG system, I connect it to the Tableau.
16:27 Nonetheless, like a JSON file, like a citation or anything, like end, end product is Tableau. So we can connect the really lots of, lots of like crazy, a lot of like input policy, ugly data source finally into a mapped and working very fine like Tableau product.
16:45 So personally, I was very like amazed by the this, um, experimental, and I feel amazing like possibility and opportunity for the Tableau developer as well. Because everyone is actually like talking about AI, and AI already give a like a possibility to leverage like data capacity. But end up, we need our ideation, right? And we need our execution. So this is the main question I want to tell you, and I want to share with you because AI has so many great technical, powerful like things, so everyone is talking about AI, and AI is literally replacing all of like data, data analyst job as well.
17:20 Not only data analyst. But at the same time, as a Tableau user and a Tableau lover, why do you still need a Tableau? And I really want to ask this question so many times, and I, I got my answer from my side. Because human perception feature are like sixty thousand faster than the text. That is actually like a scientific evidence, and I really love this one.
17:42 Because AI provide the what in general, right? We can actually get, oh, what is actually inside of this like platform or this like data source. But data is actually raw intelligence. But after we integrate AI with the Tableau, like Visual Impact, that is actually providing us a so what. So what means like you can get the next step of like a what. So what is like– So what is, what is the next, uh, what kind of insight, what is the action plan we can actually get in very intuitiv- intuitively and easily.
18:13 So that's why I believe Tableau empower this decision, especially working with AI. So That's what I want to tell you. And, uh, kind of like, so what is very meaningful, like answer from my side, so I just like decided to keep working on like Tableau and AI integration. And I believe so many of you are really like get interested how I can actually using the AI and using the Tableau, not only for using AI, not only for the like Tableau, but integration.
18:41 But this is a kind of like a sample I want to share with you. So I told you I developed a RAG system, and this is a sample RAG system I am developing for like a nonprofit like project. This is for everyone. So you already saw this dashboard so many times, but the good part of like RAG system-driven Tableau is very good. You can put your any kind of like resource, then based on your like data source, AI engine is literally giving you the strategic briefing.
19:08 Like, uh, Gemini do or like a OpenAI do, they give like… give you a very good, like briefing, right? So we can get, oh, this is like a reasoning chain based on our RAG system. I, I can follow these steps. And also we can find a micro citation as well, so like we can verify this kind of information is not based on hallucination. But at the same time, we can always go back and we can connect to our dashboard and our evidence and our data visualization storytelling.
19:37 Because I love long texting, I love evidence, but we at the same time need to capture something very intuitively. So I believe, uh, AI-driven Tableau is like the, is going to be the next things. Even nonetheless about like how agentive or how generative like it could be, it would be very good for us, it could be very efficient for us to use it, but at the same time, we need to know how to combine the AI concept and Tableau concept.
20:06 Not only Tableau, data visualization concept, and keep the right insight is our still role, and this is actually a good chance for us to work on it. So this is overall like, uh, my ideation that I want to share with you, especially the ones who really want to know how can I actually leverage AI and what kind of idea I can actually implement as a Tableau developer or as a data specialist.
20:28 This could be a kind of like ideation, and I would really want to recommend you to try from the basic start, especially following the Will Sutton. He already gave like a really great, very simple, easy-to-follow steps how you can actually connect your LLM in a like Tableau Cloud. And then from there you can expand your idea. You can actually develop the RAG system or your own like another version of like AI-driven chatbot or generative like dashboard as well.
20:55 I think that is what Trisan is actually doing. But still, so many like Tableau lover, like our like core users, they want to maximize how we can actually combine these two concept and then like can make a better impact on it. So that's what I want to share with you, and if you have any question, please let me know. I'm really happy to get, uh, get you answered, and I love to talk and communicate.
21:16 Thank you.
21:18 Chris McClellan: Thank you for that. Uh, Wendy has a question. I don't know, Wendy, if you want to ask your question.
21:26 Wendy Small: Uh, hi. Thank you so much for the presentation.
21:30 It's great. I just wanted to ask a quick question, going back to your pre-RAG, um, uh, explanation when you were actually, excuse me, connecting to the large language model.
21:44 Harim Jung: Yeah.
21:45 Wendy Small: Um, if we're using proprietary, um, uh, business data-
21:51 Harim Jung: Yeah
21:51 Wendy Small: …
21:52 when you're querying, um, using a MCP, et cetera, the tools within that-
21:57 Harim Jung: Yeah
21:57 Wendy Small: …
21:58 when it lands in the, uh, with the large language model, is there any, uh, data leaks through, um…
22:07 Does any of that data get, uh, used in other large language model applications like ChatGPT?
22:16 Like, is it available for that sort of application, or is it locked in?
22:21 Um, do you know what I'm… I'm just…
22:24 For example, if we were to use this solution that Will, um, so kindly shared with the data community on our own-
22:33 Harim Jung: Mm-hmm
22:33 Wendy Small: … proprietary data-
22:35 Harim Jung: Mm-hmm
22:36 Wendy Small: …
22:37 how do I know that that data isn't being u- isn't, um, you know, being used by OpenAI if your LLM is-
22:46 Harim Jung: Oh. Yeah, yeah.
22:48 Wendy Small: Yeah. Yeah.
22:48 Harim Jung: That is actually why the Tableau sales for develop like Tableau MCP. That is literally like MCP data protecting the data source. There's hundred percent guarantee like data leaking.
22:58 Wendy Small: So it's guaranteed it's not leaking.
22:59 Harim Jung: Yeah. That's why we… they develop, we are using this MCP, because we want to keep our data inside of, like, we are just using the LLM just to read it, but not leaking. Yeah. This is really important, actually. Yeah.
23:11 Wendy Small: Okay, great.
23:12 Harim Jung: Yeah.
23:13 Wendy Small: And I, I also understood from Will's, um, presentation, it's really important, uh, also when you're setting up the per- personalized tokens-
23:25 Harim Jung: Yeah
23:26 Wendy Small: … for, um, the access to Tableau Cloud-
23:30 Harim Jung: Yeah
23:30 Wendy Small: … that the permissions attached to that token, so the user, the user that, uh, where you've actually created the, the secret number or whatever it's called-
23:41 Harim Jung: Yeah, yeah
23:42 Wendy Small: … exactly, the permissions attached to that user is what you're actually giving the, um, the users of this front-end query bot-
23:54 Harim Jung: Mm-hmm. Mm-hmm
23:54 Wendy Small: … access to the same permissions.
23:56 Harim Jung: Yeah.
23:56 Wendy Small: So I think that's really important too, because, um-
24:01 Harim Jung: That's really important. Yeah. I just, like, develop-
24:03 Wendy Small: You know…
24:05 Yeah.
24:05 Harim Jung: Yeah. Yeah. Please go ahead, Wendy. Yeah.
24:07 Wendy Small: No, I was just saying that he makes a very important point, point, 'cause I've watched the video about three times now. I still don't understand-
24:14 Harim Jung: Mm-hmm
24:14 Wendy Small: … I don't understand half of it because I'm not a coder. Um, but we're lucky we've got an intern at the moment. Emmanuel's on this call as well.
24:24 Harim Jung: Mm-hmm.
24:24 Wendy Small: And he's been able to deploy Will's, um, p- Solution against the super data source just as a proof of concept.
24:33 Harim Jung: Right. Right.
24:34 Wendy Small: And I'm really excited. I want to actually deploy it against our data source, but I'm a bit nervous about, like, what happens with that data.
24:41 Harim Jung: Mm-hmm.
24:42 Wendy Small: Uh, yeah. But if it-
24:44 Harim Jung: That is-
24:44 Wendy Small: … if John's saying, yeah
24:46 Harim Jung: … that is actually a good point. So I am personally using the data source that is like a open data, like public data.
24:52 Wendy Small: Yeah.
24:52 Harim Jung: But of course, like, token and security and data leakage are, like, a major, biggest, like, cybersecurity issue with the DCIR era. And that's why regarding, like a… We are… If we actually tested, like, a small project, that is not a big deal, but, like, a really big company, they'd rather go use, like, directly the package from, like, uh, like Google or somewhere because they already provide, like, a really highly secure, like, a protected solution.
25:19 Use… When you're using LLM, not, not leaking any data and, like, it provide, uh, uh, manage all of the token very, very carefully. So that's actually like a corporate, like, side of, like, work, and we are… I don't wanna, like, downsize us, but we are more, like, testing our, our, our idea and how to combine the AI and Tableau to make a better impact is actually our role.
25:40 So yeah, I definitely… Like, thank you for your question, and, like, that, the tap for how important the security part is. I really think, yeah, resonate with your idea as well.
25:51 Okay, any other questions or?
25:55 Chris McClellan: Ilija's got a question.
25:57 Ilija Stojić: If I can get off mute, I can ask it. Hello.
26:02 Um, so regarding Tableau's MCP server, I was wondering if they had found a way yet to restrict the end user interacting with it, the one actually talking to some LLM or AI when it's doing the RAG. So the MCP has access to any number of sources, right-
26:25 Harim Jung: Mm-hmm
26:25 Ilija Stojić: … that you would want Tableau to interact with.
26:28 Harim Jung: Mm-hmm.
26:28 Ilija Stojić: But what if I, as an end user, only have access to one source? How does it prevent me from using the ChatGPT, let's say? Let's pretend I don't care about an external LLM, but how does it prevent the user from asking a question and getting a response from a, a source they wouldn't have access to in Tableau directly themselves, if that makes sense?
27:00 Harim Jung: Oh, this is really a good question. To be honest, I never thought about it because I only think about the user is, like, eagerly to ask some question regarding this one. So I will definitely, uh, you know, check this part and, like, uh, come back to you. This is really a good question. Thank you for asking. Yeah.
27:16 I see.
27:17 Ilija Stojić: Yeah.
27:18 It's m- for an actual example, like, if you have an HR dashboard, because my intent with this has been to have dashboards, but then have a widget where individuals could ask follow-up questions within the dashboard.
27:31 I think similar to what you were showing in your video, pretending there's-
27:36 Harim Jung: Uh-huh
27:38 Ilija Stojić: … salaries in an HR-
27:39 Harim Jung: Uh-huh
27:39 Ilija Stojić: … data source.
27:41 If I'm not in HR and I want to ask questions of a different dashboard, how do I prevent it from accidentally pulling up salary information-
27:52 Harim Jung: This-
27:52 Ilija Stojić: … as a user query?
27:53 Harim Jung: Yeah. That's really imp- important. So bec- like, we normally, uh, know about the concept, like, a role level security.
28:00 Ilija Stojić: Mm-hmm.
28:00 Harim Jung: And based on, like, a token, we… Based, based on our, like, a credential, it is, like, prohibited. But about, like, a chatbot, about, like, a conversation or what, that is another story. I think giving the credential from the start is also, like, a really good idea to prevent it, and, like, making another, like, role of a security about, like, uh, using the VizQL concept is actually a good point.
28:20 I, I really like this idea. I will take it. Yeah. Thank you.
28:24 Ilija Stojić: Okay.
28:24 Harim Jung: Yeah.
28:29 Chris McClellan: And just throwing in Emmanuel's question as well is, uh, are the presentation slides available? I don't usually ask that question-
28:35 Harim Jung: Yeah
28:35 Chris McClellan: … but it's an interesting question.
28:36 Harim Jung: Yeah. Sure. Yeah. Yeah.
28:38 Chris McClellan: Yep.
28:39 We'll, we'll get a copy and include it within the, within the rec- within the user group's, uh, link-
28:44 Harim Jung: Yeah
28:45 Chris McClellan: … and, and on the YouTube description as well.
28:47 Harim Jung: Yeah, sure.
28:49 Chris McClellan: Uh, all Harim's contact details are here, so LinkedIn or Tableau Public.
28:56 And we'll move on to our survey time, which is just a Mentimeter if s- if everyone wants to get their phones ready.
29:04 So scan the QR code on your phone.
29:08 And just while we're waiting for that, the… Paul had another question in the Q&A of, does it work with Tableau Server or only Tableau Cloud?
29:16 Harim Jung: Uh, I think it's only the Tableau Cloud as long as I understand, because Tableau Cloud is, like, connected with the VizQL and,
29:25 yeah, API, I think seriously, but I'm not sure about, like, a Tableau Server. Yeah.
29:31 Chris McClellan: Mm-hmm. That's okay.
29:32 Okay, we've got 13 people. That might be enough.
29:35 Okay. Um, just as an easy sim- an easy series of questions. Some of them you've never seen before. This one's an easy one, which is the same as in the chat all the time. Where do you live? Just so we've got a feeling for, um, where people are dialing into for this meeting.
29:52 Harim Jung: Oh.
29:53 California.
30:00 Chris McClellan: And I always forget that it lowercases everything, so.
30:06 Florida. Okay. It's getting pretty late in Florida, but it's pretty cool. So all throughout the US and, and East Coast Australia, which is fine.
30:14 Uh, so the next question is, what industry do you work in?
30:19 Just to show what Tableau can be used in, basically.
30:29 So a good mix, logistics, retail, insurance, automotive, mobility, legal, logistics.
30:39 A lot, right?
30:40 And the third easy question is, how many years have you been using Tableau?
30:45 Not that it matters because everyone started at Tableau at some stage having never used Tableau before.
30:52 But usually, we get a skew one side or other, and I can never tell which way it's gonna be.
30:56 That's a good mix. Okay, the trickier questions because you've never seen these before, and it, it involves a bit of typing. Do you feel more creative or analytical when it comes to using Tableau?
31:10 Everyone leans for the Andy Cotgreave answer.
31:18 So a lot of people saying it depends, but then more creative than analytical. Just.
31:24 And if Tableau could automatically do one thing for you, what would you choose?
31:31 Always good. Someone says, "Fold my washing." Okay. So then the next answer is make me coffee. Data curation, documenting LODs.
31:41 Visual design mockups. Interesting.
31:44 Data profiling warnings. Yes.
31:48 Automated tests.
31:50 Automated tests on what? Dashboard interaction or data quality, maybe.
32:01 Not, not that it's a harder question, but it obviously involves a little bit more thinking. Simple metadata extract. Yes. Which is always important when we're talking about AI as well.
32:15 Performance tuning, always interesting. Data definition.
32:20 Okay, we'll move on. We could be here forever, right?
32:23 The next question is, if you had to describe Tableau in three words, what would they be?
32:36 So pays mortgage, I think a lot of people would fall into that.
32:40 Tell me more.
32:42 Uh, this, you know, makes data sing. That's pretty cool.
32:46 Discover, discover, uncover, and act.
32:53 I think this is a problem with the word cloud. It jumps around too much. I'm trying to read the same thing that's now in a different position. Creative data outlet. All good stuff.
33:02 Pays my mortgage stands out the most. That's cool.
33:07 Okay, the next question.
33:10 Uh, what's your favorite Tableau feature?
33:17 Oh, even I've dropped out 'cause my phone won't sync up.
33:20 Oh, here we go.
33:25 Map layers, viz in tooltips, fun visibility, mapping.
33:33 Ilija says, "I ask this question in interviews." Yes.
33:39 Ilija Stojić: Twice in the last two days.
33:44 Chris McClellan: Layout containers. Noodles and meatballs. Yes. I thought noodles and meatballs were two different features, but parameters is always good. Filtering. Perfect.
33:54 Rounded corners soon. Yes, rounded corners are coming very soon, aren't they?
33:59 Okay, and the last one of these questions is, what's your favorite way to stay updated on new Tableau features?
34:15 Tableau beta program is interesting.
34:19 So you can get the new features before they even come out.
34:22 Tableau team interviews, stuff on YouTube, going to user groups and TUGs, both online and in-person TUGs.
34:30 Always good.
34:32 But that's the trick, right? Keeping up with the new features, learning what you can apply to your dashboards at work.
34:41 Awesome. So that's like our audience fast five, and now the last two questions are a little bit easier, maybe. Uh, what topics should be presented at a future TUG meeting? So at this TUG in the future.
34:54 Breaking down a Viz of the Day. Okay. And how… Oh, okay. So breaking down a Viz of the Day. If you look at the Tableau Public TUG, they do that. They get someone who's done Viz of the Day, who's got the award, and they walk through that presentation and basically present how they did it. Depends on what the person wants to show. How Vizzes of the Day get selected is a tougher question.
35:18 Um, customizing data sources into the right shape to suit the viz. Always good.
35:24 RLS implementation, so row-level security, so the users can only see part of the data source. And I think that conversation before we started talking about RLS at the same time as data source access. So two different things, possibly. Tableau Pulse, advancements in LLM applications, exactly what we've been seeing today.
35:42 Uh, so Jacob just brought up a good point in the chat. Uh, Louis Yu from Singapore does a Sec-Secrets of the viz. It's a podcast, and it's on YouTube as well.
35:54 And I'll put the link in the, the YouTube link from here of just, um, talking to people about how they built their dashboards.
36:03 You know what– As, as the title says, what's the secret behind what they did, how, how they actually build it, et cetera. Always good.
36:11 Yeah, more things like today. That's what we're here, right? It's just to look at those things you need to learn about.
36:17 And obviously, the next slide leading into that is who would you like to see present at future TUG meetings? And after TC, there's sort of a big list of people that we can, uh, contact to do that.
36:28 Or if there's anyone else that you want to see present. I'm still talking to Andy Cotgreave about his follow-up AMA. We just need to get a, um, a Yes. I knew it would come up soon.
36:42 So Harim, obviously on today, but we can have her back again for like a part two. Tableau Tim, very interesting. Will Sutton on exactly the same subject maybe.
36:50 Kirk, Will, Steve Wexler.
36:54 I do have a standing invite to everyone who wrote Big Book of Dashboards, um, to just at any time, the whole team, and I think that's, that's the sticking point as well. How do you get the whole team on at the same time because of time zones and stuff?
37:12 You know, there's always someone asleep or someone's gonna miss out on it.
37:16 Brittany would be awesome. Brittany's looking for a job at the moment, so if anyone knows, especially in the US, that, um, if there's any positions going, Brittany is on the market and wanting to find a new position.
37:30 Amanda Makulec on the same as Steve Wexler or in different meetings with Steve.
37:35 But again, that same group of people, that'd be good. I can contact all those people and see if they can spend some time with us.
37:43 Um, Kyle's been on before.
37:45 A very early meeting, so I don't know when that was, probably May or June last year.
37:50 So when TC25 was on, Kyle and Will presented VizQL and Chill at the conference, and then they did it back here a few weeks after that. So, um, that recording's on YouTube. But yeah, always good to have Kyle back. And obviously, he's just updated his blog recently on Tableau, not quite security, but how to use Tableau using other code. So it used to only be personal access tokens.
38:16 It's now expanded into connected, connected applications. It's now expanded recently. Um, I forget what the term is. UAT? Is it a universe– unified access token or a universal access token?
38:30 Brand-new thing that just came out. It's on his blog, um, explaining how to use that.
38:36 Okay, cool. Awesome intro. Thank you for that. Next TUG is around January twenty-second. Sorry, it's all in January twenty-second, just depends on where you are in the world. Uh, Tore and Steve talking about templating their dashboards and how you actually do that to just make your dashboard creation faster, more standardized, et cetera.
38:58 So you've got the QR code there, or just go back to the user group website. Um, it's all the same ti- it's all the same meeting. It just depends on where you are in the world as to when, when that meeting is.
39:09 But this recording and every other TUG online that gets recorded is on the YouTube channel at Team Data Fam, or there's a audio podcast on Spotify and Apple, so you can listen to it if you're not sitting near a computer.
39:22 And email me at team.DataFam@Gmail.com if you've got any suggestions or anything else that you need to get back to me on.
39:30 Uh, so thank you, Harim, for that presentation today. Uh, awesome presentation. Lots of questions apart from that. We'll get the slide deck and add it into the recording as well. And, um, yeah. Thank you.
39:42 Attendee: Thank you.
39:43 Ilija Stojić: Thanks, Chris.
39:43 Chris McClellan: Thank you. Good, good to see you, Ilija.
39:45 Ilija Stojić: Take care, everybody.
39:47 Chris McClellan: Awesome. Thanks a lot. Talk to you soon.