Description
AI agents need more than just prompts - they work with tools and data. This presentation explores the evolution of agentic workflows and the specific mechanisms they use to interact with databases. We’ll focus on the Model Context Protocol (MCP) and how it connects AI to Postgres, moving beyond basic RAG to more advanced data driven tools.
While RAG (Retrieval-Augmented Generation) is well-understood, it is often a one-way street. The next frontier is agents allowing models to interact with data, query, and reason over databases dynamically. The Model Context Protocol (MCP) is an industry standard that solves the fragmentation of tool-calling. This talk provides a practical, "how-to" guide for the popular open-source tools in that field and how they interact with PostgreSQL.
Bio
Gleb, a Cloud Advocate at Google, specializes in database technologies and AI integration within data-driven applications. His expertise includes relational databases, data transformation and ETL, application development, data replication and integration solutions. He is a runner completing 2-4 marathons per year. In the past he participated in expeditions dedicated to oceanic research in the Pacific as oceanologist.
Length
60 min
Transcript
1 00:00:01.820 --> 00:00:18.310 Gabor Szabo: So, hello and welcome to this presentation, to the Code Maven channel, if you are watching the video. My name, Gates Gabor. I run these events. Thank you for everyone who arrived to this presentation, live presentation. It's always fun to have
2 00:00:18.310 --> 00:00:26.479 Gabor Szabo: audience in the presentation, and thank you very much, Gleb, for agreeing to give this presentation. In a second, I'll have the
3 00:00:26.480 --> 00:00:33.579 Gabor Szabo: I'll give you the microphone, or the stage, or whatever, to introduce yourself and give the presentation.
4 00:00:33.580 --> 00:00:54.070 Gabor Szabo: Those people who are… if you're watching this in YouTube, then please start by liking the video and following the channel, and remember, below the video, you will find links to notes about the presentation, and links to further future events, where you can follow them and see what might be interesting for you.
5 00:00:54.940 --> 00:01:07.530 Gabor Szabo: I think that's enough for the first introduction, and Glebi will have the chance to introduce yourself and talk about the subject, so thank you very much again for…
6 00:01:07.780 --> 00:01:23.620 Gabor Szabo: coming. Oh, one more thing, those people who are here, that's the perk of being in the live presentation, you can ask questions, and we'll connect them, I guess, and once in a while, Gleb will answer the questions he likes.
7 00:01:23.970 --> 00:01:24.930 Gabor Szabo: Okay?
8 00:01:25.160 --> 00:01:26.190 Gleb Otochkin: Right.
9 00:01:27.260 --> 00:01:41.980 Gleb Otochkin: I'm not going to answer the questions I don't like, or that… no, no, of course not. All right. So, my name is Gleb. We are starting, to talk about agents, MCP, and Postgres.
10 00:01:42.040 --> 00:01:49.500 Gleb Otochkin: database, how we connect to each other, what agent, what MCP, and so on. My name is Gleb.
11 00:01:49.690 --> 00:02:04.380 Gleb Otochkin: I live in Ottawa, Canada. I used to be a scientist in astronaut, participating in some Pacific research expeditions, sometimes spending, like, 6 months in the sea.
12 00:02:04.380 --> 00:02:12.630 Gleb Otochkin: But, latest switch to software engineering and database management. So, I do a lot of running.
13 00:02:12.810 --> 00:02:22.180 Gleb Otochkin: I did a couple of marathons this year, Boston and Ottawa, and next month, I think this October…
14 00:02:22.290 --> 00:02:30.020 Gleb Otochkin: 10th, I'm running Lisbon, so, if you are in Lisbon, let's grab a coffee next month.
15 00:02:30.870 --> 00:02:42.970 Gleb Otochkin: I work at Google as a cloud advocate for databases. What it means, I'm kind of providing the bridge between developers' community and product team inside Google, trying to
16 00:02:43.020 --> 00:03:00.010 Gleb Otochkin: from one side, deliver the message from developer community about the product, and from other side, deliver some kind of best way how to use product to developer community. That's my role, and I think it is a lot of fun, doing that with people.
17 00:03:00.890 --> 00:03:16.219 Gleb Otochkin: So, there are some agenda from high level. We are going to talk about what agents, what MCP, how it works, then we talk about how they are connected to each other. This is how agents use Postgres, and then I'm…
18 00:03:16.220 --> 00:03:24.449 Gleb Otochkin: I want to talk about some tools, like MCP Toolbox, how it works, and what the benefit of the tool,
19 00:03:24.920 --> 00:03:42.799 Gleb Otochkin: Also, I would like to touch a little bit about remote MCP, what we have in the cloud, at Google Cloud, and we talk about NL to SQL, natural language to SQL, and hopefully, if time permits, I will show some simple demo, how it works in real life.
20 00:03:42.900 --> 00:03:51.510 Gleb Otochkin: Based on, agent, MCP, Toolbox, and Postgres database, alright? So, let's start.
21 00:03:52.960 --> 00:03:55.849 Gleb Otochkin: And the first part is,
22 00:03:56.090 --> 00:04:05.550 Gleb Otochkin: I probably… you already know what the agent is, but I still need to touch base. And the agent is… it is kind of…
23 00:04:05.890 --> 00:04:06.890 Gleb Otochkin: AI?
24 00:04:07.580 --> 00:04:12.850 Gleb Otochkin: driven application. It means the application is managed
25 00:04:13.130 --> 00:04:19.740 Gleb Otochkin: and reasoning by AI model, and it has some tools in possession.
26 00:04:19.940 --> 00:04:33.310 Gleb Otochkin: and disposal to apply some actions to real world. In case of real world, what I mean is the tool can, for example, do something on your computer.
27 00:04:33.410 --> 00:04:48.340 Gleb Otochkin: tool can do something on remote computer, and tool do… can do some action. So, technically, we have key components. It is model, tools, and inside agent, you have orchestration, which is agent brain, with
28 00:04:48.390 --> 00:04:57.390 Gleb Otochkin: Goals, instructions, memory, short-term, long-term, and everything else, with guardrail and so on. That's the agent.
29 00:04:57.880 --> 00:05:00.570 Gleb Otochkin: So, how Agent works in…
30 00:05:00.740 --> 00:05:12.760 Gleb Otochkin: very simple case. You have an agent which has in possession, for example, a reasoning tool, and acting tool. And, for example, we have agent which has,
31 00:05:13.070 --> 00:05:21.980 Gleb Otochkin: Access to weather services, And it has access to some kind of alarm signal for…
32 00:05:22.150 --> 00:05:32.990 Gleb Otochkin: user. In that case, what I am asking, set my alarm when the weather is right for running. I don't specify exactly what the weather is right for running.
33 00:05:33.540 --> 00:05:37.810 Gleb Otochkin: And I don't specify exactly where I am.
34 00:05:37.970 --> 00:05:52.060 Gleb Otochkin: But, assuming that agent keeps in memory that I am in Ottawa, and it has access to weather in Otolera, and maybe it has some kind of information about my habits and everything else.
35 00:05:52.240 --> 00:06:05.470 Gleb Otochkin: I probably have very good chance to get proper actions at the end. So, agent will check the weather in Ottawa. When it is sunny and agent likes.
36 00:06:05.470 --> 00:06:17.019 Gleb Otochkin: Agent knows that I like to run in sunny weather, and then it put alarm, okay, it is sunny outside in Ottawa, go for a run. That's the agent work.
37 00:06:17.330 --> 00:06:36.799 Gleb Otochkin: That is probably a silly and very simplified example, but it is what exactly agent is. It uses the model to go through workflow, define how it is going to execute the tools to achieve the goal. That's the agent in natural.
38 00:06:38.270 --> 00:06:44.180 Gleb Otochkin: The problem comes when we have multiple agents, And multiple tools.
39 00:06:44.420 --> 00:06:52.950 Gleb Otochkin: And we had… I was helping to work with some different implementations, And sometimes you have…
40 00:06:53.450 --> 00:07:01.739 Gleb Otochkin: tens of agents. Sometimes you have hundreds of agents. You have tens or hundreds of tools. And the problem is.
41 00:07:01.850 --> 00:07:19.790 Gleb Otochkin: The tool and agents can be developed by different team, by different framework, by different languages, and everything else, and it means for each agent, you have to define some kind of adapter, how agent going to call that particular tool, right?
42 00:07:19.790 --> 00:07:39.210 Gleb Otochkin: And, for example, how to call the weather services. You can create functions in your Perl or Python language code, or Go language code, how to call API of the weather service, but that tool will be only for that particular
43 00:07:39.280 --> 00:07:41.390 Gleb Otochkin: particular agents.
44 00:07:41.610 --> 00:07:58.429 Gleb Otochkin: When you develop it outside of agent and define some kind of protocol, how you connect to that tool, it means you have to create the same adapter for other agents as well. So, that creates tons of different adapters and tons of different implementations.
45 00:08:00.030 --> 00:08:07.179 Gleb Otochkin: It is a problem with scaling and interoperability. And…
46 00:08:07.280 --> 00:08:21.700 Gleb Otochkin: Anthropic, I believe it was Anthropic, came the first with idea. Let's put some kind of standards to that, and let's create an open protocol with standards how applications provide context to LLM.
47 00:08:22.090 --> 00:08:32.340 Gleb Otochkin: And… It is how model context protocol was invented. And, of course, everybody was Decided, oh, that's…
48 00:08:32.570 --> 00:08:43.289 Gleb Otochkin: Great idea, let's embrace it. And we, as a company, Google, and other companies created some kind of body which is trying to
49 00:08:44.570 --> 00:08:58.970 Gleb Otochkin: community-based, open-source community to try to define the standards, and keep the standards, and update the standards based on latest information, developers, development, and everything else.
50 00:08:59.100 --> 00:09:10.649 Gleb Otochkin: And you can read about everything about that. It is modelcontextProtocol.ai, where all the information, all the definition of the standards are.
51 00:09:11.560 --> 00:09:12.990 Gleb Otochkin: So,
52 00:09:13.240 --> 00:09:25.990 Gleb Otochkin: that, in essence, what we replaced. We replaced the A multiplied by T problem to more manageable A plus T problem, which is
53 00:09:26.280 --> 00:09:29.119 Gleb Otochkin: Much more simple to solve.
54 00:09:29.510 --> 00:09:35.539 Gleb Otochkin: And… the MCP itself developed over the years.
55 00:09:35.940 --> 00:09:45.569 Gleb Otochkin: So, the old classic MCP until this summer was stateful, based on JSON RPC,
56 00:09:45.740 --> 00:09:57.100 Gleb Otochkin: two zero messages, and in a nutshell, what you have is you have MCP host, where you place that MCP client, the protocol implementation.
57 00:09:57.430 --> 00:09:59.909 Gleb Otochkin: Then, you use a transport
58 00:10:00.230 --> 00:10:13.270 Gleb Otochkin: standard AO or HTTPSC to get the MCP server environment, which has different resources, tools, prompts, and everything else in possession, which
59 00:10:13.410 --> 00:10:16.700 Gleb Otochkin: can connect to external systems. So.
60 00:10:17.140 --> 00:10:25.020 Gleb Otochkin: it helps, you know how to send the request to MCP server, you know how to get in response.
61 00:10:25.130 --> 00:10:31.129 Gleb Otochkin: And the session ID is persistent during your interaction with MCP.
62 00:10:31.280 --> 00:10:44.960 Gleb Otochkin: So, essentially, what you do, the phase one, it is you discover what kind of tools you have, you load that tools to your memory, and then you use that tools to call the tools with parameters
63 00:10:44.960 --> 00:10:53.309 Gleb Otochkin: to get that inform… provide that information to MCP, and MCP server will execute tools with your parameter.
64 00:10:53.310 --> 00:10:56.249 Gleb Otochkin: And then get a result to you.
65 00:10:56.270 --> 00:11:02.759 Gleb Otochkin: During the execution, it can request additional information from you, and in that case.
66 00:11:02.870 --> 00:11:09.310 Gleb Otochkin: It is two-way communication. It is not only one way you get your HTTP post.
67 00:11:09.720 --> 00:11:20.900 Gleb Otochkin: pass that information, and that's it. No. You are getting in response, acknowledgement, and probably request for another information, piece of information, you pass again, and then you get a result.
68 00:11:21.370 --> 00:11:22.629 Gleb Otochkin: That is great.
69 00:11:22.750 --> 00:11:25.419 Gleb Otochkin: And… but it is not scalable.
70 00:11:25.690 --> 00:11:35.880 Gleb Otochkin: If you have… very high loaded, like, hundreds of agents and tens of MCP servers behind the scenes.
71 00:11:37.000 --> 00:11:41.080 Gleb Otochkin: And you want to be it… Scalable enough
72 00:11:41.210 --> 00:11:44.500 Gleb Otochkin: to increase number of ports for MCP servers.
73 00:11:45.250 --> 00:11:50.590 Gleb Otochkin: and place it behind the load balancer, it's not gonna work, really.
74 00:11:50.940 --> 00:11:54.120 Gleb Otochkin: And it is one of the reasons why,
75 00:11:54.510 --> 00:12:00.080 Gleb Otochkin: new MCP standards were adopted recently, it is this summer.
76 00:12:00.250 --> 00:12:03.319 Gleb Otochkin: Now it is stateless NCP client.
77 00:12:03.530 --> 00:12:07.459 Gleb Otochkin: and stateless replica of MCP servers.
78 00:12:07.980 --> 00:12:26.740 Gleb Otochkin: it can work behind the load balancer. What is replace? So, when you pass information to MCP Server, you actually put additional information underscore meta, where you put all information about client, some,
79 00:12:26.830 --> 00:12:31.040 Gleb Otochkin: housekeeping information for your post and everything else, but each
80 00:12:31.190 --> 00:12:40.860 Gleb Otochkin: communication. You post, you get results, it's self-sufficient, so you don't need to keep that sticky session anymore. And
81 00:12:41.080 --> 00:12:45.460 Gleb Otochkin: As a matter of fact, session ID was dropped from the standard itself.
82 00:12:45.660 --> 00:12:48.390 Gleb Otochkin: So, that was what changed.
83 00:12:48.730 --> 00:13:06.149 Gleb Otochkin: And of course, phase one, where discovery and edge cache capabilities, is completely optional for communication, because it might be already cached somewhere, for example, on edge stuff, all the tools and everything else, or it is
84 00:13:06.860 --> 00:13:25.200 Gleb Otochkin: you can just download that Prakash tools and use that tool. You know what kind of requests to pass, you know how to communicate and everything else. So, that's the main replacement. Of course, we… if you go to the site, you can read much more,
85 00:13:25.270 --> 00:13:33.870 Gleb Otochkin: Detailed description, what exactly has changed during the communication, and what exactly is on the roadmap.
86 00:13:34.130 --> 00:13:39.679 Gleb Otochkin: So, that's how MCP, in general, works.
87 00:13:41.670 --> 00:13:49.189 Gleb Otochkin: Couple of… few… really few, practical, notes about MCP itself.
88 00:13:50.500 --> 00:13:58.990 Gleb Otochkin: the MCP in general, what I understood from my working… from working with different companies and different developers.
89 00:13:59.460 --> 00:14:18.539 Gleb Otochkin: Yes, MCP can be some kind of resemblance of API, but in reality, what you probably want to build tools around your critical user journey. So, for example, if user is asking, I want to get a report
90 00:14:18.710 --> 00:14:20.869 Gleb Otochkin: For the last month.
91 00:14:21.050 --> 00:14:32.150 Gleb Otochkin: And you know the users are asking that particular case report for last month, for that particular piece of your information, or sales, or anything else.
92 00:14:33.170 --> 00:14:41.139 Gleb Otochkin: What you want in that case, you want to create the tool which will be executed for that intent.
93 00:14:41.650 --> 00:14:48.369 Gleb Otochkin: And… you don't want to provide, okay, you want to report for last month. Here, the…
94 00:14:48.530 --> 00:14:50.939 Gleb Otochkin: schema. Here are the…
95 00:14:51.070 --> 00:15:02.139 Gleb Otochkin: tools like execute SQL statement, check the metadata, and everything else, then you probably can execute several tools and get reports last month.
96 00:15:02.430 --> 00:15:13.819 Gleb Otochkin: That's maybe not the most optimal way, and the reason is, in that case, you spend much more time, and spent much more talking as well.
97 00:15:14.440 --> 00:15:21.089 Gleb Otochkin: So, when you build tools, think about goals and what kind of outcome you want to get.
98 00:15:21.190 --> 00:15:29.390 Gleb Otochkin: The arguments itself, as well, less arguments better for any kind of tool, and
99 00:15:29.490 --> 00:15:42.289 Gleb Otochkin: what we… what at least I see in my experience, when you have multi-layer dictionary-type argument, it doesn't really work very well with the model.
100 00:15:43.100 --> 00:15:46.390 Gleb Otochkin: Then, name tools appropriately.
101 00:15:46.760 --> 00:16:00.480 Gleb Otochkin: Keep in mind, we are still working with large language models. It means they understand language. If you name your tool drop database, but in reality tool is executing SQL statement.
102 00:16:00.790 --> 00:16:08.610 Gleb Otochkin: universal SQL statement. That probably kind of, not intuitive and
103 00:16:08.810 --> 00:16:12.410 Gleb Otochkin: Agent might not pick up that tool to execute the statement.
104 00:16:12.870 --> 00:16:17.680 Gleb Otochkin: But it will try to pick up that tool to drop database.
105 00:16:18.010 --> 00:16:23.630 Gleb Otochkin: If you ask to drop database, and in both cases, it might be not successful.
106 00:16:24.530 --> 00:16:33.979 Gleb Otochkin: Think about context. Every text output log is going to context. Of course, we try to cache context as much as possible, but still.
107 00:16:34.330 --> 00:16:42.410 Gleb Otochkin: More description for your tool. If you decide, I have a tool, and I have a several-page description, what tool does.
108 00:16:43.100 --> 00:16:53.439 Gleb Otochkin: And if you have 10 of such tools in your MCP server, it will clog your contacts and might reduce quantity of execution.
109 00:16:53.950 --> 00:16:58.189 Gleb Otochkin: So, again, and MCP is not API.
110 00:16:58.490 --> 00:17:08.590 Gleb Otochkin: don't make your MCP server exactly resembling your API access to your database management system, to your data, or anything else.
111 00:17:08.700 --> 00:17:23.650 Gleb Otochkin: In some cases, it is justified, but there are only few such cases. I maybe touch-based on some cases later in the presentation, but that's the general advice. It doesn't cover everything, but at least something, right?
112 00:17:24.569 --> 00:17:25.540 Gleb Otochkin: So…
113 00:17:25.770 --> 00:17:33.949 Gleb Otochkin: I'm stopped here, and if we have any kind of questions, let me, let me know. If you…
114 00:17:34.710 --> 00:17:37.159 Gleb Otochkin: If you are with me, let me know as well.
115 00:17:39.890 --> 00:17:42.970 Gabor Szabo: So far, I didn't see any questions. I…
116 00:17:44.810 --> 00:17:49.349 Gabor Szabo: If anyone wants to ask questions, then please either write it in the chat.
117 00:17:49.590 --> 00:17:56.709 Gabor Szabo: Or raise your hand, and then probably I… actually, I can allow you now to unmute yourself if you prefer that one.
118 00:18:02.450 --> 00:18:03.200 Gleb Otochkin: Alright.
119 00:18:03.560 --> 00:18:12.380 Gleb Otochkin: So if we don't have any questions for now, just… if you come up with the questions later, after next chapter, I will stop again, so if you…
120 00:18:12.620 --> 00:18:21.190 Gleb Otochkin: Still thinking and, have some questions, you're welcome to ask. Okay, let me continue,
121 00:18:22.270 --> 00:18:25.560 Gleb Otochkin: to talk about how the agents and Postgres
122 00:18:25.750 --> 00:18:27.709 Gleb Otochkin: Can be connected to each other.
123 00:18:29.820 --> 00:18:30.930 Gleb Otochkin: So, first.
124 00:18:31.450 --> 00:18:41.139 Gleb Otochkin: Postgres database. I know probably you already know what it is, but for people who are not familiar with different database engines.
125 00:18:41.450 --> 00:18:49.830 Gleb Otochkin: It is just a couple of words about Postgres. It is a relational database, it means, it is
126 00:18:50.050 --> 00:18:58.729 Gleb Otochkin: Supports relational data schema when the different data are connected to each other.
127 00:18:59.640 --> 00:19:04.910 Gleb Otochkin: The example… examples of different relational database, it is…
128 00:19:05.180 --> 00:19:11.519 Gleb Otochkin: for example, Oracle, one of the old databases, it is Microsoft SQL Server.
129 00:19:11.520 --> 00:19:36.370 Gleb Otochkin: MySQL, all those databases are relational databases, and in reality, relational database, it is like a Swiss knife of the database. It can run reports for you, it can use different joints to connect different tables to get proper information, and it can support all TP workload, when all TP workload, it is when you
130 00:19:36.370 --> 00:19:41.809 Gleb Otochkin: do a lot of small changes to the database. Think about that, like,
131 00:19:41.980 --> 00:19:47.940 Gleb Otochkin: Backend database for online transaction system for your…
132 00:19:48.580 --> 00:20:03.969 Gleb Otochkin: workshop, or for your store online system. So, that's the Postgres database. It is not as young as you think. In reality, it was originated in 1986,
133 00:20:05.520 --> 00:20:22.930 Gleb Otochkin: in Berkeley, and open source release was in 1997. And, of course, it is not Postgres. When it was originated in 1986, it was Ingress, and if you think about it, now it is 2026,
134 00:20:23.210 --> 00:20:28.920 Gleb Otochkin: You can calculate by yourself, it is 40 years ago, right? So…
135 00:20:29.490 --> 00:20:35.880 Gleb Otochkin: It is really a robust, reliable, data, database.
136 00:20:35.880 --> 00:20:56.049 Gleb Otochkin: with all support for ACID, integrity, consistency, and everything else. And it is highly, highly extensible, so you can create extension for Postgres database and add a new functionality, like PGVector, to work with the vector data type with embeddings in Postgres.
137 00:20:56.350 --> 00:21:10.700 Gleb Otochkin: So, Google has tons of different databases, and we support Postgres or Postgres-compatible database. Cloud SQL for Postgres, essentially the same Postgres as you have in,
138 00:21:10.820 --> 00:21:17.690 Gleb Otochkin: in the community edition, but it is just running on Google infrastructure with management layer around that.
139 00:21:17.800 --> 00:21:24.880 Gleb Otochkin: AloDB fully Postgres-compatible database. It means you have a Postgres database.
140 00:21:25.030 --> 00:21:38.369 Gleb Otochkin: and you want to move to AloDB, you don't need to change any code, but AlloDB has additional features which probably might be helpful or might not be helpful to you, depending on your workload.
141 00:21:38.370 --> 00:21:50.080 Gleb Otochkin: AI integration, columnar engine, and some performance improvements, for highly intensive workload and transactions.
142 00:21:50.170 --> 00:22:01.820 Gleb Otochkin: And we have some Postgres interface for Spaniel database. It is not real Postgres, but it is kind of adapter interface for some Postgres-like queries and actions.
143 00:22:02.260 --> 00:22:03.250 Gleb Otochkin: So…
144 00:22:03.680 --> 00:22:18.220 Gleb Otochkin: how agents use database. The first one of the cases, I think it is quite interesting, and I saw several implementations of that. The database is used for persistent memory of agent.
145 00:22:19.010 --> 00:22:36.409 Gleb Otochkin: The second is, I saw prompt management, resided inside database. It is when the database, is storing different versions of prompt, and they, probably system instructions, everything else, and then used
146 00:22:36.590 --> 00:22:42.440 Gleb Otochkin: By the agent to improve or evaluate some different versions of the prompt.
147 00:22:42.600 --> 00:22:49.599 Gleb Otochkin: Rug. Rug stands for, Retrieval Augmented Generation, when the…
148 00:22:49.800 --> 00:23:00.090 Gleb Otochkin: Agent is using information, proprietary or domain-specific information from the database to improve the answer and ground it
149 00:23:00.120 --> 00:23:11.650 Gleb Otochkin: Not based only on open source search or other publicly available information, but also on your proprietary internal information.
150 00:23:12.810 --> 00:23:15.690 Gleb Otochkin: But from other side, Agent Ken.
151 00:23:16.020 --> 00:23:31.440 Gleb Otochkin: be used to manage your database implementation. If your database environment has proper API, and you have developed the tools, MCP server, which can do that to your database. For example, taking backups.
152 00:23:31.840 --> 00:23:46.140 Gleb Otochkin: or check the performance, maybe scale up or scale down, or anything like that. You can use Agent to troubleshoot your database, performance, queries, and everything else, and of course, for monitoring.
153 00:23:46.740 --> 00:23:49.619 Gleb Otochkin: So, let's talk a little bit more about that.
154 00:23:49.870 --> 00:23:59.270 Gleb Otochkin: cases, the persistent memory, in agent. For example, you have daily long interaction with the agent, which
155 00:23:59.510 --> 00:24:03.580 Gleb Otochkin: comes… by days. In reality, Agent
156 00:24:04.750 --> 00:24:13.330 Gleb Otochkin: Tries to keep everything in memory, but different agent frameworks also using some small database of file system storage
157 00:24:13.390 --> 00:24:31.709 Gleb Otochkin: to offload memory from the context, because context is limited for the models, and you have to summarize it from time to time. The compressing context, it is taking a summary of existing context and discarding everything out.
158 00:24:31.910 --> 00:24:36.300 Gleb Otochkin: So, what's the problem with that? The problem is…
159 00:24:36.400 --> 00:24:49.699 Gleb Otochkin: If it is locally deployed, for example, you deploy it on Kubernetes port, your agent, and that works pretty well, you store that information, for example, on,
160 00:24:49.840 --> 00:25:04.249 Gleb Otochkin: file system or SQLite database, which deployed with the agent, but if the port is crushed, and you replace by another port, that information is lost, right? So, what you can do, you can upload it to
161 00:25:04.360 --> 00:25:12.420 Gleb Otochkin: Postgres database, for example, to the tables, and then use Semantic or hybrid search.
162 00:25:12.720 --> 00:25:25.500 Gleb Otochkin: to access that information when user… for example, user is just simple, very silly example. User is asking, hey, do you remember from agent, do you remember,
163 00:25:25.670 --> 00:25:44.309 Gleb Otochkin: I told you 3 weeks ago about my wife's birthday, and about my ideas for the gift. Can you please recall it and tell me if it is still on the sale as it used to be on that time?
164 00:25:44.810 --> 00:26:03.169 Gleb Otochkin: Technically, what you can do in the agent code, you can specify, okay, then go back, use vector search for why gift among the chance of conversation, then retrieve that information, provide to the model context, and then agent knows what the
165 00:26:03.480 --> 00:26:06.619 Gleb Otochkin: talk is about. So that's the idea.
166 00:26:06.850 --> 00:26:17.759 Gleb Otochkin: Also, you can upload the procedural, procedural, part of the, agent. For example, like, rules,
167 00:26:17.840 --> 00:26:31.790 Gleb Otochkin: guardrails, workflows, configuration, in terms of you can store it as JSON data inside the agent, and then use it when the agent loads up to the memory.
168 00:26:31.970 --> 00:26:33.709 Gleb Otochkin: So that's the idea.
169 00:26:34.650 --> 00:26:36.960 Gleb Otochkin: The prompt management, it was…
170 00:26:37.850 --> 00:26:53.439 Gleb Otochkin: It is interesting, I saw at least 3 different cases when people decide, okay, we want to improve our prompt, but we want to improve our prompt based on scientific approach. We don't want anecdotal data. Oh, looks like that prompt provides better results.
171 00:26:53.460 --> 00:27:03.810 Gleb Otochkin: We want to have versioning of the prompt and system instruction inside our database, then execute those prompts
172 00:27:03.850 --> 00:27:11.550 Gleb Otochkin: With our evaluation criterias for hundreds of times, and get appropriate average results
173 00:27:11.650 --> 00:27:20.700 Gleb Otochkin: whether the new versions of prompt is better than old versions of prompt, and so on and so forth. So that actually works really well.
174 00:27:21.070 --> 00:27:30.049 Gleb Otochkin: At least in those cases when people were introducing me to those systems. It was quite interesting.
175 00:27:30.780 --> 00:27:55.299 Gleb Otochkin: I have to mention the rack, it is Retrieval Augmented Generation. It boils down when you have, for example, some kind of proprietary knowledge, your knowledge base inside Postgres database, you create… you split it to the chunks, for example, you put the vector embeddings for each chunk, and then what you do, the consumer, one of the colleagues is asking the system, for example, can you maybe
176 00:27:55.400 --> 00:28:14.609 Gleb Otochkin: provide me a solution for that particular problem. And then, what agent does, agent is, okay, let me check… I'm not going to check in the internet, because you don't have that information in internet. It is your internal problem. It is your internal solution. So, I'm connecting to the database.
177 00:28:14.640 --> 00:28:30.149 Gleb Otochkin: I'm using, for example, vector search, I'm checking that information, I'm retrieving that solution back to my context, then passing context to the model, and then getting the grounded
178 00:28:30.230 --> 00:28:38.739 Gleb Otochkin: best response from the model itself, which is based on your internal proprietary data. That is… Rug in natural.
179 00:28:41.190 --> 00:28:49.510 Gleb Otochkin: So, when you do it on the scale, Sometimes you need to combine Different types of search.
180 00:28:50.010 --> 00:29:07.280 Gleb Otochkin: Plus you want to apply some index on your vector search, for example, or non-vector search, to make it really fast and scalable. And again, we are talking about relational database, and in relational database, you can
181 00:29:07.620 --> 00:29:17.859 Gleb Otochkin: use different filters joins And hybrid search, and data search, and all together to provide the best and most
182 00:29:18.200 --> 00:29:23.769 Gleb Otochkin: performance response. So, It is why
183 00:29:23.990 --> 00:29:40.519 Gleb Otochkin: I think… I can be biased here, of course, and I am probably biased, because I like Postgres. I think Postgres, as a universal database engine, suits very well for such kind of workload.
184 00:29:40.810 --> 00:29:42.770 Gleb Otochkin: You can use…
185 00:29:43.140 --> 00:30:01.979 Gleb Otochkin: consolidated data architecture. You can use all the data types, all variety of the data types, JSON, JSON, B, relational standard data types, like character numbers and everything else, plus additional vector data type, or full-text search, all together.
186 00:30:02.060 --> 00:30:12.300 Gleb Otochkin: to get… retrieve the results relatively quickly. Quick enough for the agent, because what I understand when you work with the agent.
187 00:30:12.930 --> 00:30:17.610 Gleb Otochkin: The database part retrieval of that information is really quick.
188 00:30:18.120 --> 00:30:32.679 Gleb Otochkin: part with the model is taking more time, usually. It is not the most critical part, the database, from retrieval. It is when the people, oh, the Postgres will retrieve my data in 50 milliseconds.
189 00:30:32.830 --> 00:30:38.659 Gleb Otochkin: But if I use… for example, specialized vector database, and can…
190 00:30:38.980 --> 00:30:48.140 Gleb Otochkin: it's limited down to 25 milliseconds, right? And then you wait for 2 seconds to get response from the model. So it is…
191 00:30:48.330 --> 00:30:51.340 Gleb Otochkin: Of course, again, it is definite in each case.
192 00:30:51.620 --> 00:31:02.549 Gleb Otochkin: And another important part, if you have different parts of your information in different databases, for example, vector search and bind database.
193 00:31:02.550 --> 00:31:16.590 Gleb Otochkin: some other type of information, another database. Essentially, you need to combine all those sources together to provide the answer. In that case, you need to create some kind of complicated pipeline to connect them each other, and
194 00:31:16.770 --> 00:31:25.899 Gleb Otochkin: Make it consistent that vector search and the information from non-vector database are consistent among themselves.
195 00:31:26.010 --> 00:31:32.790 Gleb Otochkin: So that creates some kind of problem. The Postgres database is always consistent. Consistent.
196 00:31:33.930 --> 00:31:45.140 Gleb Otochkin: Conversational analytics for database, it is an interesting case when you create an agent which accepts natural language input and
197 00:31:45.490 --> 00:31:59.240 Gleb Otochkin: provide back either a SQL statement, or execute that SQL statement directly inside database, and then provide the results. So… and I will try to show some short demo about that.
198 00:31:59.360 --> 00:32:06.299 Gleb Otochkin: Very basic, but still probably valuable to show the different ways how to do that.
199 00:32:08.830 --> 00:32:19.359 Gleb Otochkin: And management databases. It is already here, it is already here in… some different agents. For example.
200 00:32:19.990 --> 00:32:31.380 Gleb Otochkin: At Google Cloud, you already have ability to manage your databases in MCP servers, because it has that MCP tools available for you. For example, to need
201 00:32:31.990 --> 00:32:47.999 Gleb Otochkin: start, stop your instance, create new instance, create database, for example, I want to backup my database, and MCP Server, if it has the tool, and if you have proper permissions to execute that tool on the backend database.
202 00:32:48.040 --> 00:32:58.229 Gleb Otochkin: It will do it for you. So, it is already here, it is already available. So, it is not something we are talking about tomorrow, it is here, it is now.
203 00:32:59.340 --> 00:33:04.499 Gleb Otochkin: So, I'm stopping here, and if we have any questions, I'm happy to answer.
204 00:33:31.810 --> 00:33:34.080 Gleb Otochkin: Any questions, guys?
205 00:33:39.750 --> 00:33:41.420 Gleb Otochkin: Am I still connecting?
206 00:33:41.740 --> 00:33:45.769 Gabor Szabo: Yeah, yeah, yeah, you're, you're, you're connected, and we hear you.
207 00:33:46.920 --> 00:33:53.479 Gabor Szabo: Either people are overwhelmed, Or you are just answering everything that they are… they would ask.
208 00:33:53.530 --> 00:33:55.610 Gleb Otochkin: Maybe I'm so good.
209 00:33:55.930 --> 00:33:56.670 Gabor Szabo: Yeah.
210 00:33:57.060 --> 00:34:06.369 Gleb Otochkin: I doubt about that, but… okay, if we don't have any questions, let's go forward, how we… about the time?
211 00:34:06.370 --> 00:34:12.419 Gabor Szabo: There was just one comment that says, no questions yet from me, presentation really great and clear so far.
212 00:34:12.920 --> 00:34:14.130 Gleb Otochkin: Great, alright.
213 00:34:14.130 --> 00:34:17.399 Gabor Szabo: There's an expectation that maybe it's not… that's it.
214 00:34:17.409 --> 00:34:27.469 Gleb Otochkin: Right. Okay, let me talk about, MCP toolbox.
215 00:34:27.979 --> 00:34:35.609 Gleb Otochkin: So, we were talking about databases, how they use MCP, what is MCP, what agent. So.
216 00:34:36.059 --> 00:34:40.089 Gleb Otochkin: what kind of tools we have. And here, again,
217 00:34:41.259 --> 00:34:50.299 Gleb Otochkin: I'm always biased, everybody is biased, but our team at Google created a tool, it is an open source tool, it calls MCP Toolbox.
218 00:34:50.659 --> 00:34:58.389 Gleb Otochkin: for databases. It supports Google Database and non-Google database, Equally.
219 00:34:58.549 --> 00:35:07.329 Gleb Otochkin: It means, if you have… Cloud SQL, yes, you can use MCT Toolbox for database, but if you have
220 00:35:08.119 --> 00:35:25.729 Gleb Otochkin: open source Postgres, or MySQL, MySQLite, or Valky, or Elasticsearch, or Apache, Cassandra, or Oracle, you can use MCP Toolbox as well. It is open source, it is available, it has API,
221 00:35:26.109 --> 00:35:39.959 Gleb Otochkin: we have a great team behind the development of MCP Toolbox I'm working closely with. They are publishing a lot of interesting blogs and everything else. It has enhanced authorization support.
222 00:35:40.269 --> 00:35:49.059 Gleb Otochkin: It has SDK, and SDK for goal language, or Python, I believe for Java as well, so…
223 00:35:49.349 --> 00:35:58.419 Gleb Otochkin: You can develop something just using SDK for Toolbox and Toolbox, together. So, that's a great tool.
224 00:35:58.789 --> 00:36:03.639 Gleb Otochkin: And… If you prefer to work with IDEs.
225 00:36:03.879 --> 00:36:13.599 Gleb Otochkin: then, again, you can just… since it is MCP, it is very easy to implement in any kind of tools supporting MCP servers.
226 00:36:13.959 --> 00:36:23.689 Gleb Otochkin: Technically, for example, for… here's the example of configuration for local development. You put the path to MCP toolbox.
227 00:36:23.699 --> 00:36:34.849 Gleb Otochkin: And then you put the arguments how you want the MCP toolbox to start. Here, the argument, it is pre-built for Postgres and standard I.O, it means it is developing
228 00:36:34.939 --> 00:36:52.919 Gleb Otochkin: next to my tool, right? It is, for example, I have Xcode, or I have anti-gravity tool, and I'm just providing that information to my anti-gravity tool. Here, the MCP server you can use, and then you provide Postgres host port database user password.
229 00:36:52.979 --> 00:37:02.249 Gleb Otochkin: And in my case, for example, it is all local database and everything else, and then you are working with database out of box. That's the idea.
230 00:37:02.569 --> 00:37:13.909 Gleb Otochkin: And speaking about the pre-built tools, it means it is universal, some kind of predefined default set of tools you can use with Postgres, but
231 00:37:14.039 --> 00:37:15.709 Gleb Otochkin: Doesn't mean that it's the…
232 00:37:15.829 --> 00:37:32.919 Gleb Otochkin: effective for, for example, for user application, but for development, pre-built tools, probably the most efficient way how to use, because you don't really know, sometimes, what kind of critical user journey you will have when you develop something.
233 00:37:33.469 --> 00:37:46.739 Gleb Otochkin: But also what you can, you can provide the tools and configuration file, and exactly set of tools and tool sets you want to use with your agent as well. And I will show the example later about that.
234 00:37:47.249 --> 00:37:48.209 Gleb Otochkin: So…
235 00:37:48.549 --> 00:37:55.239 Gleb Otochkin: Here, the GitHub repository. It is, again, open source. You go to GitHub repa, you get it to your
236 00:37:55.379 --> 00:38:07.769 Gleb Otochkin: environment, and you use it. And we have documentation, mcptoolbox.dev site, so have a look at how it suits for you, what kind of database engine you use, maybe
237 00:38:07.949 --> 00:38:11.429 Gleb Otochkin: Give it a try, and let us know how it works.
238 00:38:13.589 --> 00:38:22.249 Gleb Otochkin: It is… MCV Toolbox is great for local development, and not only local development for small environments, but
239 00:38:22.369 --> 00:38:31.329 Gleb Otochkin: If you… work… at Google Data Cloud. It is Google Cloud with Google Data Services.
240 00:38:31.479 --> 00:38:40.849 Gleb Otochkin: and you want something more enterprise-ready, you might to look to MCP at Google Cloud. So, essentially what it is.
241 00:38:41.029 --> 00:38:45.009 Gleb Otochkin: It is Google Managed MCP Server.
242 00:38:45.349 --> 00:39:04.709 Gleb Otochkin: And it is MCP Server for different database engines. For example, BigQuery, MCP Server for LoadB, MCP Server for Cloud SQL, MCP Server for other data services. This unified interface, the same protocol for BigQuery, Database Looker, or anything else, is fully managed.
243 00:39:04.799 --> 00:39:10.139 Gleb Otochkin: You don't need to worry about how you're gonna scale it behind the scenes.
244 00:39:10.269 --> 00:39:21.659 Gleb Otochkin: And, built specifically for the reasoning partners on LLM. So, Technically, what you do, you…
245 00:39:22.159 --> 00:39:30.229 Gleb Otochkin: configure MCP server, and then you get a number of tools. And do you remember when I was talking that MCP
246 00:39:30.569 --> 00:39:37.119 Gleb Otochkin: tools are not API. So, here's the case where MCP tool
247 00:39:37.429 --> 00:39:42.009 Gleb Otochkin: might resemble API access, and the reason behind that
248 00:39:42.159 --> 00:39:53.509 Gleb Otochkin: For example, we at Google Cloud, we don't really know what kind of user journey you will have as a customer, right? And when you connect
249 00:39:53.619 --> 00:40:00.899 Gleb Otochkin: We cannot create, as of now, a custom MCP tool on Google Cloud MCP Server for you.
250 00:40:01.769 --> 00:40:17.749 Gleb Otochkin: So, what we provide, we provide some kind of universal tool. Okay, you can, for example, clone instance, you can start instance, you create backup, it is about management your environment at Google Cloud. Also, you can execute
251 00:40:17.919 --> 00:40:37.269 Gleb Otochkin: SQL statement, you can troubleshoot your SQL query, and some… so we provide some basic user journey for you, and then you combine that in your application with intent and everything else, and how they combine different tools together. So that is how it works. Also, it has
252 00:40:37.399 --> 00:40:45.269 Gleb Otochkin: internal data protection. So, apart from the… all the permissions you have to have at Google Cloud.
253 00:40:45.569 --> 00:40:57.879 Gleb Otochkin: it also integrated with model armor, for example, and model armor will help you to prevent you from prompt injection. By the way, that is a real thing. If you tell to
254 00:40:58.189 --> 00:41:03.429 Gleb Otochkin: For example, to your agent, do not
255 00:41:04.289 --> 00:41:11.629 Gleb Otochkin: select from user's table. It is just a wild example. I don't want users' data to be linked.
256 00:41:12.039 --> 00:41:19.979 Gleb Otochkin: It doesn't mean agent will always, refuse to get… Data from users table.
257 00:41:20.369 --> 00:41:24.229 Gleb Otochkin: It means you have to protect it by some more reliable way.
258 00:41:24.509 --> 00:41:26.809 Gleb Otochkin: The prompt injection to get,
259 00:41:27.049 --> 00:41:31.209 Gleb Otochkin: Around such kind of simple guardrail is…
260 00:41:31.529 --> 00:41:40.079 Gleb Otochkin: quite simple. You can ask, for example, any model, create me prompt injection, and it will create it for you.
261 00:41:40.289 --> 00:41:48.739 Gleb Otochkin: Of course, maybe now we have some more guardrail, but again, it is not the big deal to work around such stuff.
262 00:41:49.019 --> 00:41:50.159 Gleb Otochkin: In the…
263 00:41:50.389 --> 00:42:09.589 Gleb Otochkin: MCP Toolbox, you can create also some special… you can start MCP Server with, for example, read-only access to the database, and you can provide special configuration options to prevent write access to the database as well. So, you have different options as well.
264 00:42:10.479 --> 00:42:19.389 Gleb Otochkin: So, it is enabled by default, but it doesn't mean everybody can connect and start using it. You have to provide AM,
265 00:42:19.609 --> 00:42:23.269 Gleb Otochkin: permissions to execute MCP tools.
266 00:42:23.839 --> 00:42:31.899 Gleb Otochkin: And then you have to provide additional permissions to execute different actions on your environment. So, it is not…
267 00:42:32.119 --> 00:42:37.459 Gleb Otochkin: open for everyone, you have to explicitly provide those permissions.
268 00:42:37.569 --> 00:42:56.189 Gleb Otochkin: Your configuration of client is quite simple, it is HTTP URL, it is, for example, for Cloud SQL, it is SQLadmin, Google APIs, com, MCP, and then you provide your Google credentials, you use Google SDK to provide OAuth, or you use API key. So, that's how it works.
269 00:42:57.459 --> 00:43:04.839 Gleb Otochkin: And then you can create table, load, database, schema, run SQL, everything out of box. So… That's…
270 00:43:05.049 --> 00:43:13.289 Gleb Otochkin: MCP server, tools, and what you can have out of box. Google Cloud, and I'm… oh?
271 00:43:15.939 --> 00:43:17.399 Gleb Otochkin: What is happening?
272 00:43:19.529 --> 00:43:20.419 Gleb Otochkin: Oh.
273 00:43:24.069 --> 00:43:26.499 Gleb Otochkin: Sorry, guys, I think it was…
274 00:43:27.429 --> 00:43:35.279 Gleb Otochkin: Gemini decided to do something about, me. All right, natural language…
275 00:43:35.280 --> 00:43:36.660 Gabor Szabo: prompting the injection.
276 00:43:36.880 --> 00:43:43.269 Gleb Otochkin: Yes, it was clear prompt injection. I didn't expect that, for sure. It was not planned.
277 00:43:44.310 --> 00:43:45.250 Gleb Otochkin: So…
278 00:43:45.930 --> 00:44:00.740 Gleb Otochkin: what we have for NL2SQL, of course, I will show a couple of examples NL2SQL a little bit later, but what we developed at Google Cloud, and I think you should try that. It is kind of cool. So.
279 00:44:01.460 --> 00:44:10.509 Gleb Otochkin: By default, when you use Agent and MCP with access to your database, you technically can ask Agent.
280 00:44:11.220 --> 00:44:13.749 Gleb Otochkin: Give me the data, and agents…
281 00:44:14.820 --> 00:44:36.550 Gleb Otochkin: models are smart enough to understand, oh, I have Postgres database, I have access to information schema. Let me check what tables I have. Okay, what's the table's name, what the table's name, columns, and everything else, and technically it can connect to each other and eventually give you the right SQL query and results from the database.
282 00:44:36.550 --> 00:44:37.540 Gleb Otochkin: That works.
283 00:44:37.950 --> 00:44:53.200 Gleb Otochkin: But it is not stable, it takes too much time, especially the first time when it is executed, when it is scanning your information schema and everything else. And next time you ask the same question, you might get different SQL query.
284 00:44:53.310 --> 00:45:08.070 Gleb Otochkin: and that SQL query can be wrong, or a SQL query can be not efficient. For example, you know what SQL query should be executed for that particular intent, for that particular type of equation from the user.
285 00:45:08.420 --> 00:45:13.019 Gleb Otochkin: what you can do with Google Cloud, you can create query data can accept
286 00:45:13.560 --> 00:45:16.499 Gleb Otochkin: And providing query data CAX set.
287 00:45:16.660 --> 00:45:22.870 Gleb Otochkin: Query template, query facets, some additional pieces of information, which helps
288 00:45:23.500 --> 00:45:27.219 Gleb Otochkin: the API on backend, it is…
289 00:45:27.410 --> 00:45:33.510 Gleb Otochkin: On API, it is another whole conversational analytic agent, which
290 00:45:33.710 --> 00:45:43.349 Gleb Otochkin: connect to each other and understand, oh, I have intent from the customer, and I know what query template will serve for that. Also.
291 00:45:43.350 --> 00:45:54.569 Gleb Otochkin: to add to that query template, I add query facet, and the facet, it is a small piece of query, for example, some kind of conditions you want to use.
292 00:45:54.600 --> 00:46:02.790 Gleb Otochkin: And how it is written will depend whether, for example, index is going to be used or not. Function-based index.
293 00:46:02.840 --> 00:46:09.740 Gleb Otochkin: for example, will be used in Postgres only if the condition is written a special way.
294 00:46:09.810 --> 00:46:18.680 Gleb Otochkin: the same way as indexes created, right? And you have create query facets for that condition, and then you combine with query template.
295 00:46:18.700 --> 00:46:29.210 Gleb Otochkin: And then you generate the query, and execute query optionally, and get results back. That's how query data context set works. And you are…
296 00:46:29.220 --> 00:46:40.030 Gleb Otochkin: able to test it, of course, at Google Cloud, we have CodeLabs, how to use with query data context, and it is available, in CodeLab, I believe.
297 00:46:40.270 --> 00:46:43.380 Gleb Otochkin: It is kind of…
298 00:46:44.090 --> 00:46:56.450 Gleb Otochkin: you have two options. You can use it with MCP Toolbox, and you can use it with Google Remote Query Data Interface at Google Cloud. So, there are two different ways.
299 00:46:59.760 --> 00:47:08.779 Gleb Otochkin: So, that is recorded DMI. I don't want to use recorded DMI. Before going… hold on, let me, go…
300 00:47:09.490 --> 00:47:12.050 Gleb Otochkin: Before going forward.
301 00:47:12.330 --> 00:47:22.520 Gleb Otochkin: Let's make sense what we have. For example, you have anti-gravity or any kind of CLI, AI…
302 00:47:22.750 --> 00:47:23.760 Gleb Otochkin: ED?
303 00:47:24.850 --> 00:47:36.589 Gleb Otochkin: you can use it with managed MCP Server as a, for example, database management person, like DBA, DevOps, or anybody else. Or,
304 00:47:36.700 --> 00:47:39.200 Gleb Otochkin: conversational analytics.
305 00:47:39.570 --> 00:47:44.449 Gleb Otochkin: With MCP Toolbox, it is more suitable for local development.
306 00:47:44.550 --> 00:47:46.969 Gleb Otochkin: And self-managed databases.
307 00:47:49.430 --> 00:48:05.469 Gleb Otochkin: when you use anti-gravity, not CLI, but anti-gravity, it is probably the best for wipe coding. And MCP Toolbox, it is vibecoding with your local database. And of course, you also have AI Studio, it is
308 00:48:05.660 --> 00:48:23.219 Gleb Otochkin: I think a great thing about AI Studio. Now, AI Studio is integrated with some database engines as well, like Firestore and Cloud SQL as well. So, that's kind of trying to make sense from all that zoo of the tools and MCP servers available for you.
309 00:48:24.760 --> 00:48:42.260 Gleb Otochkin: We have, of course, at Google, we created, some MCP resources for you, if you want to scan what, what it is there. It is, on GitHub, at Google MCP, and CodeLab, I mentioned already, for… about NL,
310 00:48:42.260 --> 00:48:44.810 Gleb Otochkin: to SQL using query data. So.
311 00:48:44.810 --> 00:48:48.799 Gleb Otochkin: Try it, try it out, let us know, and be…
312 00:48:49.040 --> 00:48:54.410 Gleb Otochkin: bef- and now, before I'm going to the demo, I would like to…
313 00:48:54.510 --> 00:49:00.410 Gleb Otochkin: stop, and if you have any questions, I'm happy to answer. Or we can do it after the demo as well.
314 00:49:06.760 --> 00:49:08.579 Gabor Szabo: I think you can do the demo.
315 00:49:08.870 --> 00:49:12.830 Gabor Szabo: And, a few people have questions, Yeah.
316 00:49:13.240 --> 00:49:15.549 Gleb Otochkin: Okay, let me go to the demo.
317 00:49:17.550 --> 00:49:22.079 Gleb Otochkin: So, speaking about the demo, what I have here, I have a box
318 00:49:23.020 --> 00:49:30.150 Gleb Otochkin: On that box, I have a toolbox, MCP toolbox, deployed here, and I have
319 00:49:30.290 --> 00:49:38.289 Gleb Otochkin: tools, it is parameter file for MCP toolbox, it is tools.yaml. Let me make it a little bit bigger, probably, right?
320 00:49:39.120 --> 00:49:40.050 Gleb Otochkin: Alright.
321 00:49:41.040 --> 00:49:42.400 Gleb Otochkin: Is it…
322 00:49:42.400 --> 00:49:42.970 Gabor Szabo: Yeah.
323 00:49:42.970 --> 00:49:44.090 Gleb Otochkin: good enough?
324 00:49:44.090 --> 00:49:48.620 Gabor Szabo: Maybe a little bit larger, I mean, I have a big screen, but
325 00:49:49.520 --> 00:49:52.840 Gleb Otochkin: Okay, fuck yeah.
326 00:49:53.820 --> 00:49:56.760 Gabor Szabo: And just… just the font's a little bit larger.
327 00:49:56.760 --> 00:49:57.470 Gleb Otochkin: Okay, it is…
328 00:49:57.470 --> 00:49:58.240 Gabor Szabo: Both of them.
329 00:49:58.240 --> 00:49:59.280 Gleb Otochkin: I'm trying to deliver.
330 00:49:59.280 --> 00:50:00.100 Gabor Szabo: No, please.
331 00:50:00.970 --> 00:50:01.500 Gleb Otochkin: Yup.
332 00:50:01.870 --> 00:50:02.700 Gleb Otochkin: Okay.
333 00:50:12.770 --> 00:50:16.599 Gleb Otochkin: why it is disconnected from other sessions, I don't know.
334 00:50:19.890 --> 00:50:22.900 Gabor Szabo: The younger people in the audience say that they can see finally.
335 00:50:25.110 --> 00:50:29.750 Gleb Otochkin: Oh, I don't know why it is disconnecting. I don't like it.
336 00:50:31.970 --> 00:50:32.910 Gleb Otochkin: Okay.
337 00:50:34.540 --> 00:50:35.580 Gleb Otochkin: But, yeah.
338 00:50:41.460 --> 00:50:42.720 Gleb Otochkin: Alright…
339 00:50:42.720 --> 00:50:45.420 Gabor Szabo: Yeah, it's much better, I don't know, yeah.
340 00:50:45.620 --> 00:50:51.380 Gleb Otochkin: Okay, let me… just to… it is 3 different windows here,
341 00:50:51.690 --> 00:50:58.910 Gleb Otochkin: 3 different windows, and those windows will be using… Now, what is going on?
342 00:51:00.530 --> 00:51:01.930 Gleb Otochkin: Yeah, okay.
343 00:51:02.330 --> 00:51:08.790 Gleb Otochkin: So… What I have here, let me clear it here, okay…
344 00:51:09.850 --> 00:51:14.090 Gleb Otochkin: I have, deployed some simple agent code.
345 00:51:14.090 --> 00:51:18.850 Gabor Szabo: I just want to say that it's a nice touch that you use Database Maven as the… Several.
346 00:51:19.810 --> 00:51:24.010 Gleb Otochkin: I created that environment for that particular presentation, yes.
347 00:51:24.010 --> 00:51:25.319 Gabor Szabo: Nice. So…
348 00:51:25.650 --> 00:51:37.180 Gleb Otochkin: So, we have here the agent, agent using, ADK deployment, and, it is relatively simple,
349 00:51:39.640 --> 00:51:44.800 Gleb Otochkin: Python script with, Relatively simple quote here.
350 00:51:44.970 --> 00:51:56.650 Gleb Otochkin: it is using minimum, all the resources and requests and everything else, so it is very, very simplified. So, it has one tool.
351 00:51:56.840 --> 00:52:01.490 Gleb Otochkin: And that tool is connecting to MCP Toolbox.
352 00:52:01.720 --> 00:52:10.049 Gleb Otochkin: deployed locally, and it is one of functions also I implemented. It is how much tokens we spend, right?
353 00:52:10.210 --> 00:52:22.319 Gleb Otochkin: That's the, agent code. And it is very, very simple. It is created based on, just a couple of commands from ADK CLI.
354 00:52:25.180 --> 00:52:38.649 Gleb Otochkin: That's… the agent. Also, what we have here, we have tools.yaml for the toolbox, and if we get…
355 00:52:41.840 --> 00:52:44.140 Gleb Otochkin: tools.yaml.
356 00:52:44.420 --> 00:52:46.330 Gleb Otochkin: That is,
357 00:52:47.100 --> 00:53:01.519 Gleb Otochkin: configuration file for MCP Toolbox, and it provides the information about the resource, database, user password, and everything else. Of course, it is in clear text, but it doesn't mean it has to be in clear text. The MCP Toolbox
358 00:53:01.760 --> 00:53:18.119 Gleb Otochkin: has very advanced authentication and everything else, so that can be done. And what I'm here providing the tools, around the user's journey. For example, the… one of the first tools, search for hotels based on name.
359 00:53:18.180 --> 00:53:30.439 Gleb Otochkin: Then, in order to search hotels by location. For example, let's say we have some kind of travel system behind the scenes, and we are trying to work with that travel system.
360 00:53:30.660 --> 00:53:41.330 Gleb Otochkin: And, based on that, also you can book hotels or update, information about hotels, right? That's kind of tools we're providing.
361 00:53:41.640 --> 00:53:45.090 Gleb Otochkin: And, of course, we have that toolbox itself.
362 00:53:45.290 --> 00:53:52.240 Gleb Otochkin: And we have a Postgres database, and… And if we go to…
363 00:53:58.430 --> 00:54:00.580 Gleb Otochkin: Local host…
364 00:54:09.450 --> 00:54:16.510 Gleb Otochkin: I believe it is a database called ToolboxDB. Okay…
365 00:54:18.050 --> 00:54:22.950 Gleb Otochkin: If we go the, database itself… What is it?
366 00:54:24.390 --> 00:54:26.429 Gleb Otochkin: I don't see the error.
367 00:54:31.480 --> 00:54:32.630 Gleb Otochkin: Two books…
368 00:54:36.170 --> 00:54:37.680 Gleb Otochkin: Is it the right one?
369 00:54:39.800 --> 00:54:42.660 Gleb Otochkin: Toolbox, user, password, my password.
370 00:54:44.380 --> 00:54:47.949 Gleb Otochkin: Oh, yeah, it is wrong here.
371 00:54:50.650 --> 00:54:53.140 Gleb Otochkin: No, it is something toilet.
372 00:54:54.360 --> 00:54:56.420 Gleb Otochkin: connection failed.
373 00:54:56.720 --> 00:54:58.990 Gabor Szabo: He doesn't like the word localhost there.
374 00:54:59.390 --> 00:55:06.670 Gleb Otochkin: Yeah, I don't know why. Let me go and… Make a little bit cheat.
375 00:55:10.660 --> 00:55:12.080 Gleb Otochkin: Oh, hold on.
376 00:55:18.300 --> 00:55:19.630 Gleb Otochkin: Alright…
377 00:55:23.120 --> 00:55:28.789 Gleb Otochkin: And here we have database, we have 2BoxDB, let's connect to 2BoxDB.
378 00:55:29.980 --> 00:55:41.020 Gleb Otochkin: And here we have some tables, and we have tables with hotels, right? And if we select… from…
379 00:55:42.770 --> 00:55:55.649 Gleb Otochkin: we can see some different information, whether it is booked, what kind of photos we have, we can check in checkout dates and everything else when they're available. So, it is very, very simple schema, right?
380 00:55:55.950 --> 00:56:16.000 Gleb Otochkin: So, if we… execute… Alright… toolbox itself, with, our config tools YAML?
381 00:56:16.260 --> 00:56:24.010 Gleb Otochkin: Here we execute, and… For example, we can now access that MCP server
382 00:56:24.140 --> 00:56:37.210 Gleb Otochkin: for example, by just using quarrel command and getting to the backend. It is how every single MCP server is supposed to work, because if you put a post
383 00:56:37.380 --> 00:56:39.279 Gleb Otochkin: For example, right?
384 00:56:39.580 --> 00:56:41.580 Gleb Otochkin: To that one. Where is it?
385 00:56:42.570 --> 00:56:45.829 Gleb Otochkin: Let me open the demo requests.
386 00:56:48.340 --> 00:56:57.800 Gleb Otochkin: Okay… If we put, for example, Simple Coral, and we…
387 00:56:58.230 --> 00:57:05.119 Gleb Otochkin: Can we list meet tools by name? And we just filter, by one output.
388 00:57:06.280 --> 00:57:18.529 Gleb Otochkin: it will provide all our tools from configuration files, what we defined, right? That is how it's supposed to work, and you technically can get more information if you just
389 00:57:18.690 --> 00:57:35.390 Gleb Otochkin: get a little bit deeper insight, and you can always search hotel by name, it is description, it is what it does, it is the name of the hotel, and everything else. So, that information is… should be available. It is how MCP server works. You put the request, you get the result.
390 00:57:35.930 --> 00:57:40.449 Gleb Otochkin: So… You also search…
391 00:57:41.560 --> 00:57:49.129 Gleb Otochkin: you can do a little bit more advanced and everything else. So, that works really well, right? But…
392 00:57:49.520 --> 00:57:53.560 Gleb Otochkin: It is for our tools, but it doesn't mean we,
393 00:57:53.840 --> 00:58:09.410 Gleb Otochkin: have to use the same way. Do you remember about the pre-built tools? When we don't know what kind of user journey is going to be, we don't know what kind of tools we want to create. It is more suitable for development itself. So, in that case.
394 00:58:10.390 --> 00:58:17.450 Gleb Otochkin: What we can do… We can provide the…
395 00:58:18.600 --> 00:58:24.400 Gleb Otochkin: toolbox with parameter. I don't know if you can see here, it is toolbox with…
396 00:58:25.410 --> 00:58:30.520 Gleb Otochkin: pre-built Postgres tools. So, it requires some extra,
397 00:58:31.570 --> 00:58:44.070 Gleb Otochkin: environment variables to work with it, but essentially, it is the same, like, Postgres host, password, user, and everything else. So, when you start with pre-built tools.
398 00:58:44.350 --> 00:58:59.729 Gleb Otochkin: and go back and try to list the tools, what you have, tools by name, you are getting completely different set of tools. It is default tools for Postgres database. It is list stored procedures.
399 00:58:59.770 --> 00:59:16.950 Gleb Otochkin: stats, available extensions, and everything else, and you can execute query, you can get query plan, for example, and so on. So those pre-built tools are when you don't really know what kind of critical user journey you have, but you want to use that MCP server.
400 00:59:17.000 --> 00:59:21.660 Gleb Otochkin: Right? And so, it is how it is working behind the scenes. So…
401 00:59:22.110 --> 00:59:25.130 Gleb Otochkin: In that case, your agent can choose
402 00:59:25.680 --> 00:59:33.949 Gleb Otochkin: from pre-built tools and go through complicated, journey, for example, right? So…
403 00:59:34.450 --> 00:59:41.850 Gleb Otochkin: Let's… what we can try to do right now, we can… oh, white is jumping?
404 00:59:42.400 --> 00:59:45.620 Gleb Otochkin: Let's start the agent interface here.
405 00:59:46.720 --> 00:59:51.939 Gleb Otochkin: And the agent interface, it is, again, ADK,
406 00:59:52.060 --> 01:00:05.270 Gleb Otochkin: I'm starting, using UV Run ADK web, it is Python, developer. And of course, ADK has different languages. ADK, it is, Agent Development Framework from Google.
407 01:00:05.390 --> 01:00:14.419 Gleb Otochkin: So, I'm starting the web interface. What it allows me to, connect, using the web interface
408 01:00:15.320 --> 01:00:19.109 Gleb Otochkin: Let me try to and see if it works here.
409 01:00:19.260 --> 01:00:21.080 Gleb Otochkin: Yeah, it works here.
410 01:00:22.700 --> 01:00:27.890 Gleb Otochkin: So, we have a new session,
411 01:00:29.290 --> 01:00:48.650 Gleb Otochkin: And what we can do, remember, we have pre-built tools, we don't have predefined user journey. The agent, by itself, have no idea what exactly behind the database. What agent has… agent has the basic prompt, you are database assistants.
412 01:00:49.020 --> 01:00:51.439 Gleb Otochkin: For example, we put hello here.
413 01:00:51.970 --> 01:00:57.300 Gleb Otochkin: And agent know, oh, I have toolbox toolset.
414 01:00:58.160 --> 01:01:07.750 Gleb Otochkin: And I am connecting to Toolbox Toolset, and understand what I can do. And here, what it is responding. Let me make it a little bit bigger.
415 01:01:10.330 --> 01:01:13.280 Gleb Otochkin: It… it is responding, hello.
416 01:01:13.590 --> 01:01:23.010 Gleb Otochkin: how can I assist you with your database today? Feel free to ask about table inspection. So, it doesn't really know what it can do. It is universal.
417 01:01:23.170 --> 01:01:24.710 Gleb Otochkin: What I can ask.
418 01:01:27.890 --> 01:01:34.869 Gleb Otochkin: Can, you find a hotel in Basel?
419 01:01:37.670 --> 01:01:46.200 Gleb Otochkin: And let's see what it is going to do. So, since it is null, it knows I can work with Postgres database.
420 01:01:46.350 --> 01:01:54.050 Gleb Otochkin: And it is what it is doing behind the scenes. It is smart enough to list tables, understand
421 01:01:54.350 --> 01:01:57.639 Gleb Otochkin: Then, understand the table structure.
422 01:01:57.750 --> 01:02:02.010 Gleb Otochkin: and then execute SQL statement and provide me information.
423 01:02:03.200 --> 01:02:06.990 Gleb Otochkin: Only concern for me, it is how many steps
424 01:02:07.240 --> 01:02:16.649 Gleb Otochkin: Requires to provide that information, and if you can see here, total tokens, it is 36,000 tokens, right?
425 01:02:16.840 --> 01:02:18.170 Gleb Otochkin: So, that's…
426 01:02:19.320 --> 01:02:29.209 Gleb Otochkin: not big, but it is still kind of significant number of steps, so we spent 12 steps. Just to be fair, if I ask…
427 01:02:30.050 --> 01:02:32.990 Gleb Otochkin: What about hotel?
428 01:02:33.680 --> 01:02:36.840 Gleb Otochkin: in… Montreal.
429 01:02:39.250 --> 01:02:46.610 Gleb Otochkin: It will spend much less steps, it is only two steps, and essentially you have only 50,000 tokens.
430 01:02:46.940 --> 01:02:49.570 Gleb Otochkin: Two questions, which is not big, but still.
431 01:02:49.710 --> 01:02:55.779 Gleb Otochkin: We are working with very simplified database schema.
432 01:02:55.960 --> 01:03:13.129 Gleb Otochkin: it would be way more questions and back and forth if your schema has 15 different tables, and the information is spread around different tables, and it has to make a connection from one table to another, and another, and another. So, that's the idea. So.
433 01:03:14.540 --> 01:03:27.350 Gleb Otochkin: What if we do the same, but… we… start our… toolbox with our pre-configured.
434 01:03:27.530 --> 01:03:36.640 Gleb Otochkin: tools around user journey, right? Let's do that. We're starting our toolbox on MCP server, It is,
435 01:03:36.910 --> 01:03:50.569 Gleb Otochkin: with tools YAML, we are restarting our agent. Remember, we had 50,000,581, tokens so far?
436 01:03:50.860 --> 01:03:54.700 Gleb Otochkin: Let's restart our agent from scratch.
437 01:03:57.310 --> 01:04:04.970 Gleb Otochkin: put here, new session, and do exactly the same, questions. We are asking, hello.
438 01:04:05.610 --> 01:04:09.650 Gleb Otochkin: And see what it can work with.
439 01:04:11.370 --> 01:04:21.010 Gleb Otochkin: And you can see that even first response, the general hello, is already different, because it analyzed the tools and possessions.
440 01:04:21.240 --> 01:04:34.220 Gleb Otochkin: from the MCP, and oh, I can work with hotels here, and now I'm responding differently. I can work with hotel database, feel free to ask.
441 01:04:34.450 --> 01:04:42.000 Gleb Otochkin: Can you find me a hotel in Basel?
442 01:04:44.270 --> 01:04:46.929 Gleb Otochkin: And I'm asking the same question.
443 01:04:48.790 --> 01:04:53.480 Gleb Otochkin: And them getting, less… steps.
444 01:04:53.830 --> 01:04:57.419 Gleb Otochkin: I'm getting predefined, response.
445 01:04:57.900 --> 01:05:02.010 Gleb Otochkin: And… You can compare the number of tokens.
446 01:05:02.700 --> 01:05:14.370 Gleb Otochkin: This 8,000… plus 400 tokens, ex… instead of 36,000, right? Let's… What about…
447 01:05:24.320 --> 01:05:31.900 Gleb Otochkin: And the second question, it is, again, two steps. I'm getting, the information, what I want, and…
448 01:05:32.680 --> 01:05:36.199 Gleb Otochkin: The number of tokens is only 13,000.
449 01:05:37.020 --> 01:05:42.989 Gleb Otochkin: Versus, you probably remember, we had… Where is it?
450 01:05:44.130 --> 01:05:46.400 Gleb Otochkin: 50,000 tokens before.
451 01:05:46.720 --> 01:05:56.799 Gleb Otochkin: That's the difference between the tools created for user journey, specified agent for MCP, and tools created just…
452 01:05:57.370 --> 01:06:12.350 Gleb Otochkin: general purpose, and let an agent to understand how to work it out with all the metadata and everything else behind the scenes. That's my demo, and now probably I should switch back to the questions.
453 01:06:13.860 --> 01:06:21.940 Gleb Otochkin: I'm stop sharing right now, and… Let's go back to that.
454 01:06:24.170 --> 01:06:27.990 Gabor Szabo: Your, your screen is… your camera is off.
455 01:06:29.770 --> 01:06:37.659 Gleb Otochkin: aw… Okay, let me… let me find… Where is my interface?
456 01:06:39.160 --> 01:06:40.770 Gleb Otochkin: Boy, I can't…
457 01:06:40.770 --> 01:06:42.060 Gabor Szabo: It disappears.
458 01:06:43.750 --> 01:06:44.590 Gleb Otochkin: Oh.
459 01:06:46.410 --> 01:06:51.339 Gleb Otochkin: I, I, I, I'm trying to find my interface. Oh, yeah, I know where it is, yeah.
460 01:06:51.880 --> 01:07:00.000 Gleb Otochkin: Okay, I'm back. So… Any questions so far, guys?
461 01:07:00.400 --> 01:07:06.519 Gleb Otochkin: I'm happy to answer and discuss anything what you've seen so far, heard so far.
462 01:07:06.630 --> 01:07:08.149 Gleb Otochkin: What do you think about it?
463 01:07:10.580 --> 01:07:12.580 Gabor Szabo: Yeah, you… could you… oh, okay.
464 01:07:16.040 --> 01:07:19.089 Gabor Szabo: Sorry, now you can unmute yourself. I turned it off.
465 01:07:19.610 --> 01:07:27.360 Alejandro Imass: I think it was very clear. There are some things I already knew about some of these things that you explained. Obviously, you're focusing it to the Google…
466 01:07:27.930 --> 01:07:34.049 Alejandro Imass: Cloud, obviously, they… The object of the presentation,
467 01:07:34.800 --> 01:07:45.939 Alejandro Imass: But yeah, it clarified some context. I didn't know about the stateless MCP, so I learned quite a few things. Very interesting, presentation, very clear. No questions for me, it was very, very clear.
468 01:07:46.280 --> 01:07:49.049 Alejandro Imass: Very well presented, so I really liked it.
469 01:07:49.210 --> 01:07:50.460 Alejandro Imass: I think it was worthwhile.
470 01:07:51.250 --> 01:07:51.910 Gleb Otochkin: Thanks.
471 01:07:55.330 --> 01:08:01.979 Gabor Szabo: If anyone else wants to still ask questions here during the video, that's, that's fine, or say anything.
472 01:08:02.360 --> 01:08:06.499 Gabor Szabo: If not, then, then we'll, we, we'll…
473 01:08:06.960 --> 01:08:13.790 Gabor Szabo: And the video, and then those people who are here can stay on, and we can.
474 01:08:13.790 --> 01:08:20.019 Gleb Otochkin: I see some questions… I see some questions in the notes from Emmanuel,
475 01:08:20.740 --> 01:08:33.889 Gleb Otochkin: How do you control and monitor token usage when the large database schemas or query results are injected into LM context? That's a great question.
476 01:08:34.649 --> 01:08:39.600 Gleb Otochkin: you… So, the first couple of things.
477 01:08:39.750 --> 01:08:41.840 Gleb Otochkin: The token's usage
478 01:08:42.439 --> 01:08:53.249 Gleb Otochkin: depends on the context. Contents, sorry. So, it is, I had a presentation, probably, year or two…
479 01:08:53.359 --> 01:08:57.170 Gleb Otochkin: ago, how the different type of content
480 01:08:57.620 --> 01:09:04.220 Gleb Otochkin: imparts that token usage. For example, with,
481 01:09:05.910 --> 01:09:20.529 Gleb Otochkin: Content for, some large language models, you can have very big contents window, you can put entire book, War and Peace, from Leotal's story to that contents.
482 01:09:20.529 --> 01:09:27.809 Gleb Otochkin: But on the same time, you cannot put even small database table in the same contents.
483 01:09:28.000 --> 01:09:36.240 Gleb Otochkin: which is 100 or thousands times smaller in size. Why is that? Because
484 01:09:37.740 --> 01:09:41.329 Gleb Otochkin: Talking is not a symbol of worth.
485 01:09:42.109 --> 01:09:48.319 Gleb Otochkin: And token can be different. When you put the text in the tokens.
486 01:09:48.870 --> 01:09:55.949 Gleb Otochkin: Then the tokens will be recognized by either one word or part of the word.
487 01:09:56.360 --> 01:10:05.019 Gleb Otochkin: And that is great. But when you have, for example, flight table, for your United flights.
488 01:10:05.150 --> 01:10:06.820 Gleb Otochkin: with numbers.
489 01:10:07.290 --> 01:10:10.119 Gleb Otochkin: Then, each number will be talking.
490 01:10:11.460 --> 01:10:15.019 Gleb Otochkin: It means each number, and each…
491 01:10:15.800 --> 01:10:26.280 Gleb Otochkin: point, or anything else, we'll be talking as well. So, in that case, it is first thing. So, it is why you want to, limit how much,
492 01:10:26.520 --> 01:10:33.579 Gleb Otochkin: Information, how much Information you want to get out of the database.
493 01:10:33.720 --> 01:10:36.840 Gleb Otochkin: and provide to your OM context window.
494 01:10:37.540 --> 01:10:45.820 Gleb Otochkin: And you need to remember that not every content can be cached the same way on LLM site.
495 01:10:46.060 --> 01:10:53.220 Gleb Otochkin: And in your agent as well, so… If you can't cash content, on…
496 01:10:53.920 --> 01:11:00.339 Gleb Otochkin: Llm side, that's great. If your agent framework allows you to do that, that's…
497 01:11:00.680 --> 01:11:16.990 Gleb Otochkin: good for you. If you can't, then you have to slim down any… for example, during retrieval augmented generation, you don't want to retrieve whole table, you want to retrieve only few rows to provide contents. So, that's how you control it.
498 01:11:17.660 --> 01:11:20.710 Gleb Otochkin: Is it easy to do? Not really, it is hard.
499 01:11:21.210 --> 01:11:34.059 Gleb Otochkin: you know what you're… what you're working with. You have to know your data, you have to know your schema, you have to know how… what kind of tools to provide your MCP server to minimize that interaction with your data.
500 01:11:35.390 --> 01:11:41.530 Alejandro Imass: I have a… I have… I actually had a question. And it's regarding context. When you mentioned it, it reminded me…
501 01:11:43.190 --> 01:11:44.530 Alejandro Imass: I guess,
502 01:11:45.040 --> 01:11:53.590 Alejandro Imass: all LLMs, and I'm guessing all agents, will suffer from context lag as the context grows, the conversation grows over time.
503 01:11:53.860 --> 01:11:56.749 Alejandro Imass: The models get exponentially slower.
504 01:11:56.890 --> 01:12:01.319 Alejandro Imass: Like, inversely, I just want to show this lower, and pretty quickly. It doesn't take much.
505 01:12:01.920 --> 01:12:06.530 Alejandro Imass: To saturate the… so you have to reboot these agents, very frequently and clear their…
506 01:12:06.870 --> 01:12:08.789 Alejandro Imass: Contact start all over again, or…
507 01:12:09.220 --> 01:12:14.279 Alejandro Imass: What is the strategy there for the… I'm guessing the agents are the ones that are going to start lagging pretty quickly.
508 01:12:14.840 --> 01:12:17.649 Alejandro Imass: As they… as the context grows.
509 01:12:18.240 --> 01:12:22.119 Alejandro Imass: context of the conversation, a conversation we're having right now with the agent.
510 01:12:22.370 --> 01:12:26.549 Alejandro Imass: about trips. It doesn't take a lot to saturate,
511 01:12:27.190 --> 01:12:29.139 Alejandro Imass: And to start, you know,
512 01:12:30.910 --> 01:12:33.920 Alejandro Imass: Providing bad context to the conversation, and, you know.
513 01:12:34.030 --> 01:12:39.580 Alejandro Imass: hallucinations, stuff like that. So, what is the… the strategy there? Like, rebooting the agent, frequently? What…
514 01:12:39.690 --> 01:12:45.339 Alejandro Imass: I mean, you talked about scale at the very beginning, is one of the reasons, but… and one of the things that I have found
515 01:12:45.810 --> 01:12:51.450 Alejandro Imass: Is that it scaled horribly because of… and a growing conversation will,
516 01:12:51.600 --> 01:12:55.880 Alejandro Imass: and very quickly, saturate most elements out there.
517 01:12:56.890 --> 01:13:02.470 Gleb Otochkin: Yeah, good question. So, yes, longer you keep conversation with the agent.
518 01:13:03.430 --> 01:13:17.399 Gleb Otochkin: more context you have, like, historical context, all the interactions, and depending on the model, and of course, on the model side, in most modern… model as a service, it is going to be cached, right?
519 01:13:17.660 --> 01:13:19.990 Gleb Otochkin: So you don't really,
520 01:13:20.870 --> 01:13:29.810 Gleb Otochkin: do a lot of back and forth again, but still, it is getting slower and slower and slower, because even if it is cached, it has to be analyzed.
521 01:13:30.350 --> 01:13:32.490 Gleb Otochkin: For each single response.
522 01:13:32.810 --> 01:13:35.089 Alejandro Imass: And I found no easy way, there's no…
523 01:13:35.090 --> 01:13:36.279 Gleb Otochkin: So, the…
524 01:13:36.280 --> 01:13:38.740 Alejandro Imass: That way, to clear the context and start over, so it's…
525 01:13:38.760 --> 01:13:42.529 Gleb Otochkin: But, you, you can compare your contents.
526 01:13:42.800 --> 01:13:59.079 Gleb Otochkin: What I mean by other compress your contents, what you can do… do you remember we were talking about episodic memory for, on the database side? So what you can do, you can store all that long conversation on the database side, but from time to time.
527 01:13:59.200 --> 01:14:12.420 Gleb Otochkin: You… ask… Internally, your agent, create… when the context size reach certain size, create summary of that context.
528 01:14:13.140 --> 01:14:18.880 Gleb Otochkin: and replace the context. So you're creating short summary what was before.
529 01:14:19.370 --> 01:14:25.820 Gleb Otochkin: and discarding the main context body. So you make it small again, without losing,
530 01:14:25.990 --> 01:14:32.060 Gleb Otochkin: subject of conversation itself. That's… we call Campreya's context.
531 01:14:32.620 --> 01:14:37.210 Gleb Otochkin: Of course, if you don't… if you discard it and forget about that, that…
532 01:14:37.400 --> 01:14:40.279 Gleb Otochkin: Reduce your quantity of your responses.
533 01:14:41.570 --> 01:14:47.120 Gleb Otochkin: Because sometimes you have to retrieve that old memory, and that is where you want to
534 01:14:47.220 --> 01:14:53.119 Gleb Otochkin: Search inside your database and get the piece of chunks of context back to your conversation.
535 01:14:53.120 --> 01:15:04.230 Alejandro Imass: I see the… I see what you're proposing as a pattern is kind of keeping the growing context in the database, only breach it, or only access it if you need to, but try to work with the summary.
536 01:15:04.380 --> 01:15:06.030 Alejandro Imass: Yep. For most cases.
537 01:15:06.160 --> 01:15:10.340 Alejandro Imass: Okay, I got it, I got it. That's a… that's a nice pattern. Okay.
538 01:15:12.540 --> 01:15:13.430 Alejandro Imass: Wow.
539 01:15:13.430 --> 01:15:15.910 Gleb Otochkin: Another question, I believe, was…
540 01:15:20.290 --> 01:15:24.759 Alejandro Imass: It was interesting, because when Alberto told me about this workshop.
541 01:15:25.860 --> 01:15:29.839 Alejandro Imass: The idea of agents and databases didn't immediately click.
542 01:15:31.380 --> 01:15:37.030 Alejandro Imass: But now I see value in this, in this idea. Yeah.
543 01:15:37.030 --> 01:15:47.519 Gleb Otochkin: So, we actually have a lab, code lab, my colleague, at Google Engineering, created a lab how to create that
544 01:15:47.770 --> 01:16:03.980 Gleb Otochkin: upload the context. They created lab for AloIDB, but AlloyDB, essentially, Postgres behind the scenes, so you can technically get the lab, it is just sample code, what you can do, and they created the lab using not ADK, but blank chain graph.
545 01:16:04.100 --> 01:16:11.299 Gleb Otochkin: Land graph interface framework for agent, and what… what they do, they…
546 01:16:11.490 --> 01:16:16.730 Gleb Otochkin: Use the database to store in the context and everything else, and the episodic memory, and…
547 01:16:16.860 --> 01:16:29.579 Gleb Otochkin: If your agent somehow dropped again, and you can restart it and blow out memory back, so that's… they created Lab. I didn't try that lab by myself, it just released, probably, like, a week or two ago.
548 01:16:29.580 --> 01:16:32.550 Alejandro Imass: We… there was,
549 01:16:32.900 --> 01:16:38.900 Alejandro Imass: context marking the conversations would be nice as well. You can, like, mark certain context.
550 01:16:39.270 --> 01:16:42.730 Alejandro Imass: You know, like, at a certain point in the conversation, market.
551 01:16:43.030 --> 01:16:45.900 Alejandro Imass: And then relating that to an action or something that…
552 01:16:46.330 --> 01:16:51.279 Alejandro Imass: all the response that was obtained, that those techniques I've also seen are quite interesting.
553 01:16:51.610 --> 01:16:52.590 Alejandro Imass: Yeah.
554 01:16:52.590 --> 01:17:11.580 Gleb Otochkin: And technically, you can use… when you upload the context memory to the database, you create… split it to chunks, and then, create embedding letters on that sense, and then it will much… it will be much easier to find that pieces of context using vector search.
555 01:17:13.420 --> 01:17:19.279 Alejandro Imass: Very interesting. This is… Very enlightening. Thank you.
556 01:17:20.320 --> 01:17:20.950 Gleb Otochkin: Welcome.
557 01:17:22.100 --> 01:17:23.230 Gleb Otochkin: Anybody else?
558 01:17:26.030 --> 01:17:32.199 Emmanuel Thouraud: Yes, I have another one. How do you control the SQL queries generated by the MCP?
559 01:17:32.520 --> 01:17:35.840 Emmanuel Thouraud: To prevent a poorly constructive or recursive query.
560 01:17:36.390 --> 01:17:43.480 Emmanuel Thouraud: The goal is to avoid consuming excessive PostgreSQL resource or running indefinitely.
561 01:17:43.800 --> 01:17:53.970 Emmanuel Thouraud: Are we able to set some self-ground, such as query time hours, or limiting the resource, or even a kind of query validation on the MCP itself?
562 01:17:55.140 --> 01:18:01.370 Gleb Otochkin: So, it depends. It is kind of… so, MCP is a just toll call, right?
563 01:18:01.800 --> 01:18:08.369 Gleb Otochkin: So, when you execute a tool call, you can provide some guardrails.
564 01:18:08.490 --> 01:18:10.490 Gleb Otochkin: For example, for…
565 01:18:11.590 --> 01:18:21.939 Gleb Otochkin: For example, in Google Cloud MCP Server, Managed MCP Server, I'm not talking about the options for MCP Toolbox, it has its own options.
566 01:18:22.190 --> 01:18:28.679 Gleb Otochkin: How to limit, for example, execution query, how to make it read-only, how to print on some other stuff.
567 01:18:29.360 --> 01:18:39.259 Gleb Otochkin: You… either provide For example, for MCP Toolbox, I showed we have a tool
568 01:18:39.730 --> 01:18:42.970 Gleb Otochkin: For example, search hotel by city, right?
569 01:18:43.520 --> 01:18:53.910 Gleb Otochkin: In the example, and I provide the query inside the tool. That query will be executed. So, in that case, you know what kind of query is going to be executed, right?
570 01:18:54.100 --> 01:19:09.909 Gleb Otochkin: And in that case, you don't expect anything else. But if I provide only pre-built tool, you don't know what kind of query to be… will be executed. In that case, you have few options. The first option, you first
571 01:19:10.110 --> 01:19:17.500 Gleb Otochkin: You don't want to change… if you don't want to change anything in your database, you provide read-only access to your database, right?
572 01:19:17.800 --> 01:19:20.310 Gleb Otochkin: It can be done on different levels.
573 01:19:20.500 --> 01:19:22.749 Gleb Otochkin: It can be done on user level.
574 01:19:23.210 --> 01:19:39.670 Gleb Otochkin: what user can do and what user cannot do. For example, you can select from all tables for that particular schema, but you cannot change anything on those tables, right? So that's on database level. Also, what you can do
575 01:19:39.880 --> 01:19:43.920 Gleb Otochkin: You can provide Time out for the query, as well?
576 01:19:44.790 --> 01:19:51.330 Gleb Otochkin: You can tell, or if query is executed in more than 5 seconds, terminated.
577 01:19:52.510 --> 01:19:58.259 Gleb Otochkin: connection. That's… For example, for Google Cloud, manage…
578 01:19:58.600 --> 01:20:01.949 Gleb Otochkin: MCP servers, I believe it is 30 seconds.
579 01:20:02.570 --> 01:20:08.489 Gleb Otochkin: by default, If query is executing more than 30 seconds by default.
580 01:20:08.510 --> 01:20:24.390 Gleb Otochkin: then it will be terminated. It is why, for example, I believe I… I'm not sure if I provided the example why we're providing extra parameters. If you know your query will be executed more than 30 seconds, you can provide the parameters time out for the query.
581 01:20:24.430 --> 01:20:31.409 Gleb Otochkin: In that case, you know. For example, you have analytical query, you know that query is going to run 5 minutes.
582 01:20:31.490 --> 01:20:37.499 Gleb Otochkin: And you know you're gonna wait for that query, you don't want to terminate it in the middle, so…
583 01:20:37.710 --> 01:20:39.879 Gleb Otochkin: that option, but…
584 01:20:40.000 --> 01:20:50.419 Gleb Otochkin: It is… you have to technically work based on your data and your user journey, really. It is why I say that MCP Server is not just API.
585 01:20:50.880 --> 01:21:09.410 Gleb Otochkin: It is about user journey, what kind of requests, and query data is working really well in that sense. Of course, you don't have query data with open source Postgres, but at Google Cloud, if you use query data, in that case, it has internal mechanism preventing a lot of things behind the scenes to just
586 01:21:09.500 --> 01:21:11.060 Gleb Otochkin: plain safe.
587 01:21:11.330 --> 01:21:19.559 Gleb Otochkin: Some of them are documented, some of them are not documented, but in general, it is, like, if query is executed on query data.
588 01:21:19.860 --> 01:21:25.969 Gleb Otochkin: If it will be executed in read-only by default, unless you specify it explicitly.
589 01:21:27.840 --> 01:21:31.329 Gleb Otochkin: So, that's where you can kind of…
590 01:21:31.560 --> 01:21:44.559 Gleb Otochkin: be more or less on the safe side. But again, yes, there are a lot of uncertainty, and the models are unpredictable, as you know, right? And you want to reduce that,
591 01:21:44.900 --> 01:21:47.279 Gleb Otochkin: Uncertainty and,
592 01:21:48.010 --> 01:21:55.259 Gleb Otochkin: Weigh how the model works, providing the templates, facets, and everything else to make the query what you expect.
593 01:21:58.070 --> 01:22:06.310 Gleb Otochkin: Because, yeah, it can create absolutely ridiculous query with, tons of mergers and everything else, yeah.
594 01:22:08.800 --> 01:22:09.790 Emmanuel Thouraud: Okay, thank you.
595 01:22:11.260 --> 01:22:13.529 Gleb Otochkin: I hope I answered the question, at least I tried.
596 01:22:19.230 --> 01:22:22.649 Gleb Otochkin: Any other questions, guys? We still have 3 minutes, right?
597 01:22:28.300 --> 01:22:31.020 Emmanuel Thouraud: So, yeah, so one more I have in mind.
598 01:22:32.210 --> 01:22:40.769 Emmanuel Thouraud: I think, yeah, as you spent a lot of time on this MCP, did you benchmark the approach MCP, versus a non-MCP implementation?
599 01:22:41.400 --> 01:22:41.740 Gleb Otochkin: Yeah.
600 01:22:41.740 --> 01:22:44.600 Emmanuel Thouraud: To compare the quality on, okay.
601 01:22:44.600 --> 01:22:52.939 Gleb Otochkin: So… The fir- most of comparison lately is coming from, do we need MCPU or skills?
602 01:22:53.670 --> 01:22:54.490 Emmanuel Thouraud: Yeah.
603 01:22:55.090 --> 01:22:58.010 Gleb Otochkin: So… what I found…
604 01:22:59.140 --> 01:23:08.820 Gleb Otochkin: It depends, again, it depends. When you have MCP with predefined query, usually what you compare in that case, it is, like, pre-built tools for, or
605 01:23:09.230 --> 01:23:18.800 Gleb Otochkin: like… Widely scoped MCP tools, like, default MCP tools for Postgres, and set of skills, right?
606 01:23:19.130 --> 01:23:25.050 Gleb Otochkin: Because, technically, what you do When you execute skills.
607 01:23:25.820 --> 01:23:30.240 Gleb Otochkin: Skills, it is kind of reaction to agent what to do, right?
608 01:23:30.300 --> 01:23:49.919 Gleb Otochkin: In that case, you provide direction to the agent, okay, here are the tools you have for skills. If user asks you to create database, execute that SDK call, right? And you have SDK inside your agent, and skill, okay, I have that tool to execute that SDK code.
609 01:23:49.920 --> 01:23:58.899 Gleb Otochkin: call. It can be actually a REST API call to interface, right? And that is create. Mcp will work
610 01:23:59.660 --> 01:24:03.539 Gleb Otochkin: differently, because in that case, you communicate to MCP,
611 01:24:04.410 --> 01:24:11.620 Gleb Otochkin: with the tools and parameters for the tool, and MCP will execute tools. So, what I found is
612 01:24:12.990 --> 01:24:28.950 Gleb Otochkin: sometimes, depending on type of the request, skills can be better than MCP, sometimes MCP can be better than skills, it depends on the request, but I found if you really know what your users are going to do, you combine both together.
613 01:24:30.630 --> 01:24:40.599 Gleb Otochkin: In that case, for example, sometimes it is much easier to use one tool to create something which MCP will be using later.
614 01:24:40.600 --> 01:24:41.100 Emmanuel Thouraud: there.
615 01:24:43.510 --> 01:24:48.620 Gleb Otochkin: Skills are more universal than MCP, by default, as from, you know…
616 01:24:50.500 --> 01:24:52.830 Emmanuel Thouraud: Yeah, it's sometimes more simple to set up.
617 01:24:53.290 --> 01:25:09.240 Gleb Otochkin: Yes, it is, like, you don't need to do anything just to read SkillsMD file, and everything is alright, but again, when you work with databases, it can be kind of a little bit misleading, because you might get not results you really expect.
618 01:25:09.310 --> 01:25:18.860 Gleb Otochkin: It is working really well with different SDK, but the databases You still need data interface, Right?
619 01:25:18.960 --> 01:25:35.899 Gleb Otochkin: Or you have to connect directly to your database somehow, and sometimes it is not as easy as, and as protected. And another whole, absolutely different world, I'm going to speak in a month on PGConf in New York.
620 01:25:36.130 --> 01:25:38.280 Gleb Otochkin: It is security and MCP.
621 01:25:39.890 --> 01:25:51.109 Gleb Otochkin: nobody asked here, but it is a kind of big problem with security. It is a lot of patterns you want to prevent from prompt injection, SQL injection,
622 01:25:51.510 --> 01:25:59.129 Gleb Otochkin: I mean, drop in your database, right? If you put in the prompt, don't drop my database, it doesn't mean agent will not do that.
623 01:26:00.370 --> 01:26:01.529 Emmanuel Thouraud: Yeah, of course.
624 01:26:01.700 --> 01:26:02.230 Gleb Otochkin: that.
625 01:26:07.820 --> 01:26:17.179 Alejandro Imass: REST, as well, for access. There's other… not necessarily direct connection database, we also have PGREST, PulseREST, there's a bunch of, like, REST interfaces to…
626 01:26:17.770 --> 01:26:21.700 Alejandro Imass: Postgres, I run… In the database itself, and serve as a…
627 01:26:21.990 --> 01:26:23.930 Alejandro Imass: REST API to the outside world.
628 01:26:25.150 --> 01:26:38.119 Gleb Otochkin: Yeah, it is different, different layers. Also, the layer, for example, if you want to protect only piece of your data in your schema. And in that case, you probably want to
629 01:26:39.050 --> 01:26:45.429 Gleb Otochkin: Use some database, Mechanism to prevent reading from that table.
630 01:26:45.970 --> 01:26:53.729 Gleb Otochkin: And there are some different ways. We created private security views, for that. It means,
631 01:26:53.920 --> 01:27:06.510 Gleb Otochkin: In that case, only if you are probably… it is kind of role-level access rules, which helps to prevent some data. You want to see… you need to see only what you…
632 01:27:07.290 --> 01:27:13.250 Gleb Otochkin: allows… a lot to see, right? You don't want to provide everything to all users.
633 01:27:15.760 --> 01:27:17.760 Alejandro Imass: A lot of food for thought, definitely.
634 01:27:20.660 --> 01:27:23.230 Alejandro Imass: Oh, pretty cool. I have a hard stop right now, so I have to…
635 01:27:23.230 --> 01:27:23.750 Gabor Szabo: joke.
636 01:27:23.980 --> 01:27:28.969 Alejandro Imass: Thank you very much, guys. This is a very, clip, and, Gabor for putting this together.
637 01:27:29.500 --> 01:27:32.019 Alejandro Imass: It was really great. I really enjoyed it.
638 01:27:32.340 --> 01:27:32.869 Alejandro Imass: Thank you.
639 01:27:32.870 --> 01:27:33.510 Gleb Otochkin: Thank you.
640 01:27:34.170 --> 01:27:34.760 Gabor Szabo: Hmm.
641 01:27:35.160 --> 01:27:38.159 Gabor Szabo: Yeah, Gleb. Thank you very much for this presentation.
642 01:27:38.710 --> 01:27:39.460 Gabor Szabo: And thank you.
643 01:27:39.460 --> 01:27:39.950 Gleb Otochkin: You're welcome.
644 01:27:39.950 --> 01:27:48.380 Gabor Szabo: Everyone, all the people who were here and asked questions, or didn't ask questions, you're still welcome.
645 01:27:48.770 --> 01:27:49.170 Gleb Otochkin: Yeah.
646 01:27:49.170 --> 01:27:53.919 Gabor Szabo: And, people who are watching the video… I think we are finishing now, right?
647 01:27:54.540 --> 01:27:55.390 Gleb Otochkin: Yeah. Yep.
648 01:27:55.390 --> 01:28:07.530 Gabor Szabo: Okay, so, people who are watching the video, please like the video, and follow the channel, and remember that below the video, you will find links to,
649 01:28:08.500 --> 01:28:24.089 Gabor Szabo: thinks about this presentation, probably maybe also the slides, I'm not sure, we'll see. Definitely to access to where you can find, grab, and to the future events. So, thank you very much, and see you in,
650 01:28:24.220 --> 01:28:25.419 Gabor Szabo: The next session.
651 01:28:25.860 --> 01:28:26.659 Alejandro Imass: Thank you very much.
652 01:28:26.660 --> 01:28:27.520 Alberto Mijares: Thank you very much.
653 01:28:27.520 --> 01:28:28.130 Gleb Otochkin: Thanks.
654 01:28:28.130 --> 01:28:29.949 Alberto Mijares: It's great, right? Really great.
655 01:28:30.470 --> 01:28:31.660 Marlene: Thank you.
656 01:28:32.660 --> 01:28:33.610 Marlene: Thank you.
