[music]Hello and welcome to today's show. Today we are talking about Agentic AI meets search using open search and MCP. Sema, how are you doing? >> Hi everyone, I'm doing great. How are you?>> I'm doing well. Thanks for joining me today. So yeah, one of the reasons I haveyou here on the show for people is because you had an excellent session at Open Search Con North America and I really wanted to uh kind of dive into some of what you guys did. So thewhole idea behind MCP, we hear MCP a lot, we hear AI a lot. Can you tell me a little bit about why MCP is important uh inthe context of just working with different technologies but then also with open search specifically >> right um so initially like last two three years we have heard a lot about LLMs and then came in agentic AI and then like there are so many LLM models and so many agentic AI for each tool like GitHub has its own Google has its own and open AI has its own. So to combine everything, it's so much easier and it just makes sense to have an MCP uh to combine all the power that we can have with the agentic AI into one thing. So that's how we are using it and um you can just have it running uh up and running in your local system as well. you can have uh integrated with the multiple AWS um services or open source specifically we can have it up and running in just minimum 10 seconds and it's so much cooler and then you can communicate with your data uh in natural language that's um that I think it's most convenient for new users and new team members it's very useful to them because it's very um because you know uh some organizations have very large uh infrastructure new people can't understand it better so I think it's easier so that we can communicate with MCP and just talk it through and understand the whole architecture uh and MCP is making it very easy for us now >> so whatis MCP >> it's a model context protocol uh so it's uh just like in the talk uh we mentioned that it's kind of like a USBC port which combines all the capabilities of all your underneath it's all LLM models. It just uh combines them into a single place so that you can communicate to multiple models at once. >> Yeah. One of the things I thought was really interesting as I was working with uh MCP for a different project. I was working on one recently with MCP and ClickHouse um was that I didn't have to be a Clickhouse expert to extract data and business insights >> while using MCP. I could ask it like you said natural language queries. So I could say, you know, what is um what is the stats on a certain thing over the last 24 hours >> andit goes and it knows what queries to run against um against the infrastructure or theuh technology and it returns those insights for me. So I could use that technically without have ever touching or logging into ClickHouse itself and still gain value. So I can see that MCP is really valuable for people that are not already experts and even for the people that are experts. It gives them the ability to quickly get more insight out of um out of the technology. >> Yeah, that's very true. And uh with agentic AI, you can actually write the code uh and uhI've seen like it if uh I mean it's just normal uh agentic AI process. So if it doesn't find the uh result that you are looking for, it just keeps on retrying and it's like okay let me think this okay this command is not correct let me rethink it and make it a script. So it's really cool to see that uh capability and uh using that as well. So it is just empowering all the engineers out there. >> Ilove it. So tell me a little bit about your demo. Let's break this down. So if I were to set the context here, uh if I remember the diagram correctly, you had um open search running, you had um some data in there, you had MCP connected, and thething that I thought was really interesting as well is you had a different front end to communicate with open search. So maybe start there and tell me why you had a different front end and what the benefit is ofdoing that. >> Yeah. Uh so the normal approach that uh open search documentation also is providing is we can use cursor or uh um clawed uh desktop UI. We can use any of that but we thought that it would be just cool to show that we have a front end and that actually is not just a front end that you can just chat with. It's not just that but we had different sections of uh the UI. So we had the first section I think I didn't show that in demo Idon't remember but there was a first section where we can see the cluster health uh for open search if it's healthy if all of the index indices are present or not and one of the cool features was that we can choose the context we want to communicate with. So um I think in organizations we can have maybe 100 plus indices and then it's really hard to communic uh just get down to the result because it will take longer for the MCP. So we thought that we will give two options either we can choose everything. Um so there is one scenario where you want to uh just explore all of your cluster all of the indices and maybe there is a correlation or just uh explore one of the index so it's easier and uh it's faster to get the results. So we had that option as well and uhother was like directly communicating u through the MCP and uh we were using OpenAI uh for that in the back end. >> So then um as we talk about this so you had your uh your front end and that was really neat because you're basically putting guard rails for anyone. you give someone a location to come to um to communicate with your open search cluster and then you technically have some guardrails where you're not giving them admin privileges into the cluster itself. You're giving them just a website where they can um chat back and forth about the uh the open search cluster.>> So one of the problems that then comes in is cost and tokens andthenumber of tokens that are being sent back and forth. What did you find from that? Did you were you able to minimize the number of tokens being used andreally make this cost effective? >> Yeah, so for Okay, so the cost is really a main thing and it was just our personal project. So we didn't want to spend any money because we are running our uh local MCP right. So we thought uh we explored a lot of options but the thing uh I did was I went to open AAI um like I think we use chat GPT4 uh 4.x text and then uh the default value thereare some charges I think $5 $10 whatever usage that you have right so we thought why even do that right uh so there is uh I just made a new account and um for that they give some uh tokens free of cost but since our model wewere just trying to test it out and how it's responding so we wanted to focus on that testing and it can take longer time and it can consume a lot of tokens as well. So we um added the code uh I think from 3 million tokens we reduce it to 60 tokens so that it just doesn't generate uh random stuff that we don't want in the results. So um we did that using just uh I think just adding the Python uh script. Yeah. And adding all the indices. So that reduced a lot because otherwise what um AI does is itsometimes just give a lot of text that we don't even want. So I think reducing that and getting the right results is really important and that's why we chose to reduce the tokens and u yeah and that did not cost at all. So yeah for personal project yes. So you said it went from 3 million down to 60 or 60,000. or 60,000. or 60,000. >> 60K. >> 60K.Got it. >> And then the Python script, is that basically just giving some parameters and telling it cut out extra words? >> Yeah.It'sjust a simple Python script. Yes. Python script. Yes. >> Okay. So basically with uh with this you're giving it some context andrules around how to respond and doing that alone saved you 200 or 2.9 million tokens. >> Yeah. Yes. >> Okay. Wow. So, if we were to take the lessons you learned from doing this um andpreparing your presentation and your demo for Open Searchcon, what are some key takeaways that you would give other businesses and people that want to try and use uh you know, Agentic AI with open search? >> Okay. Um I would say um okay, first thing is we don't really want to use MCP everywhere, right? it has to be a specific use case. If you really want to see the um the help that you can get from MCP, definitely go ahead. But itjust doesn't make sense every time applying it. But uh I would like to say that um for MCP uh standalone MCP was is really great to use for open search. And if you are already on open search 3.x text definitely you can use the SSC and HTTP protocol that has all the support and um cost uh like AI comes with a cost so you need to refine and fine-tune your models as well um and the best thing was uh we tried actually I tried with a lot of models so initially like while I was testing it out I used claude 4 first that givesamazing results. Uh but that comes with a cost. So we switched to OpenAI and that's okay. I mean if you want to spend then definitely Claude 4 was giving results like um we asked like uh if there are any 5xx or 4xx errors in my application. It gave the exact count and what was the root cause. So it was amazing to see that but the same results were not seen in OpenAI uh GPT4. So it's it depends which model you are using. Uh MCP will help you but I think the underneath is the LLM model that you're going to use. So uh make sure to test that out first. Yeah. >> Awesome. Any other words of wisdom that you want to share with uh with the community here? UmI'm not sure like what else I can share. Uh I mean open search is a great project. Uh I think just this year I explored it. Um like last year I was I just heard about it that there is something called open search and it's a vector database because all the AI things were booming up and this year I got to try a lot of features not just MCP.came in recently. uh but before that uh I also presented at other open search conferences. So uh it's a really great tool actually for anomaly detection, machine learning and uh KN&N search, semantic search. It's a great tool. It's a power pack. Everything is in just one tool. So it's really great and we can connect um open telemetry as well uh to open source. So it's great. uh you can get all the observability in just one place. >> Yeah. >> Yeah. >> Yeah. >> Awesome. Thank you so much and thank you so much for joining me. For everyone that's still watching us, thank you for being here. Hit subscribe, hit like, and we'll see you in the next video. Thanks. Bye-bye.>> Thank you. [music] Bye.