WEBVTT 00:00.000 --> 00:07.000 Thank you very much and thanks everyone for coming. 00:07.000 --> 00:11.000 That's amazing to see the full house here in the search Devroom and excited to be here 00:11.000 --> 00:17.000 at first them 2026 and on this session we'll talk about open search and search innovations 00:17.000 --> 00:19.000 with open search v3. 00:19.000 --> 00:21.000 I'm Dutant Horvets. 00:21.000 --> 00:24.000 I'm open search ambassador under the Linux Foundation. 00:24.000 --> 00:29.000 I'm also a sincere ambassador and then my day Joe by Miss Senior open source advocate at AWS. 00:29.000 --> 00:38.000 And I'll be co-speaking with Ashwatt Senior Search Engine architect for open search at AWS. 00:38.000 --> 00:44.000 So for those who don't open search it's essentially a platform for AI powered search, 00:44.000 --> 00:47.000 observability and analytics. 00:47.000 --> 00:54.000 On the search side it provides the full suite from lexical search to a dense vector search, 00:54.000 --> 01:00.000 we have the interesting talk now hybrid search, multi-model search, agentex search. 01:00.000 --> 01:03.000 So the full end to end search capabilities. 01:03.000 --> 01:07.000 It's Apache 2.0 open source license. 01:07.000 --> 01:11.000 It's a top-level project under the Linux Foundation. 01:11.000 --> 01:16.000 It started back in 2021 as a fork of elastic search and kabana. 01:16.000 --> 01:22.000 After these projects were released from Apache 2 to non-open source licensing scheme. 01:22.000 --> 01:27.000 So the community wanted to keep them open, AWS stepped up and initiated the fork. 01:27.000 --> 01:33.000 And then fast forward to 2024, AWS transferred the project to the Linux Foundation. 01:33.000 --> 01:36.000 So it's owned by the Linux Foundation or any vendor. 01:36.000 --> 01:44.000 And there's the open search software foundation under the LF to foster open governance and open collaboration. 01:44.000 --> 01:50.000 And I'm really happy to say that many companies have already joined the foundation. 01:50.000 --> 01:55.000 And you see some examples here, Uber, SAP, NetApp, IBM and others. 01:55.000 --> 02:01.000 And are taking part in studying the future of this project. 02:01.000 --> 02:09.000 This is an asset code base, 140 reports under the GitHub org, 1.6 billion downloads to date. 02:09.000 --> 02:16.000 It's a super active project, more than 3,000 active contributors from more than 400 different organizations. 02:16.000 --> 02:19.000 It actually, I think, I checked the LFX leaderboard. 02:19.000 --> 02:25.000 It's I think the 12th top active project under the LF amongst the nearly 1000 projects. 02:25.000 --> 02:28.000 So pretty impressive. 02:28.000 --> 02:34.000 And the biggest news in the past year is being that after 3 years of 2.x, 02:34.000 --> 02:38.000 we finally reached version 3, the next major release. 02:38.000 --> 02:42.000 And this major release has foundational changes. 02:42.000 --> 02:46.000 Also Apache Lucine moving to version 10. 02:46.000 --> 02:54.000 We have architectural changes, such as the reader write a separation to separate between the search and the indexing workloads. 02:54.000 --> 02:59.000 Streaming architecture, pool based ingestion, native MCP support. 02:59.000 --> 03:06.000 Many new commands on the PPL, the piped processing language, the query language we have as part of the project. 03:06.000 --> 03:11.000 And support out of the box for Apache calcite as the engine and much much more. 03:11.000 --> 03:15.000 So on this session, we'd like to share with you some of the highlights of V3. 03:15.000 --> 03:19.000 And I would like to hand it over to us to start showing us this stuff. 03:19.000 --> 03:20.000 Go ahead, that's what. 03:20.000 --> 03:21.000 Thank you. 03:27.000 --> 03:28.000 Okay. 03:28.000 --> 03:31.000 So, I'll start, thank you. 03:31.000 --> 03:34.000 The passing code is the neural spaces with an open search. 03:34.000 --> 03:37.000 I like to call this as a budget-friendly semantic search. 03:37.000 --> 03:41.000 And if you have used the dense vectors, you will probably agree with me. 03:41.000 --> 03:47.000 The neural space model that is there in hacking phase, pre-trained by open search, 03:47.000 --> 03:51.000 is one of the most popular, most downloaded space model. 03:51.000 --> 03:55.000 But the catch with the neural spaces has always been the scalability. 03:55.000 --> 03:57.000 Beyond 10 million documents. 03:57.000 --> 03:59.000 The latency is... 04:00.000 --> 04:03.000 Yeah. 04:03.000 --> 04:07.000 So, beyond 10 million documents, the latency is quite high. 04:07.000 --> 04:10.000 To an extent where I start recommending dense vectors. 04:10.000 --> 04:14.000 But at that point, we are no longer in the budget-friendly semantic search. 04:14.000 --> 04:18.000 So, with this seismic algorithm implementation, 04:18.000 --> 04:21.000 we are able to push past 10 million ceiling. 04:21.000 --> 04:25.000 We are able to go to billion with the B. 04:26.000 --> 04:30.000 And this is by being able to do approximate sparse retrieval. 04:30.000 --> 04:36.000 So, with the new data structure, the forward index that actually has the dot-hyd to the 04:36.000 --> 04:38.000 first pass vector mapping. 04:38.000 --> 04:41.000 And then we have the inverted index, which probably most of you understand, 04:41.000 --> 04:46.000 but with some slight changes, where the posting list goes through a pruning process. 04:46.000 --> 04:49.000 It only retains that top weighted dot IDs. 04:50.000 --> 04:55.000 So, for example, let's say the term, dot, exist in dot 1, dot 2, dot 5. 04:55.000 --> 04:59.000 And each document that term, dot exists with different weights. 04:59.000 --> 05:04.000 So, pruning essentially removes those dots that have lower weights. 05:04.000 --> 05:10.000 And then what we do is the group similar documents, forming a cluster. 05:10.000 --> 05:13.000 And for every cluster, as summary vector is created. 05:13.000 --> 05:16.000 So, this is what you see in block 1, block 2. 05:16.000 --> 05:19.000 So, it's basically the cluster with the summary. 05:19.000 --> 05:24.000 And now that the index is created, during search time, the query vector, 05:24.000 --> 05:26.000 queries against the summary vector. 05:26.000 --> 05:30.000 If the score is too low, the cluster is simply skipped. 05:30.000 --> 05:33.000 If the score is decent enough, then the cluster is examined. 05:33.000 --> 05:38.000 And for the full final score, we go to the forward index, test the full 05:38.000 --> 05:41.000 speed to do the final score calculation. 05:41.000 --> 05:43.000 So, this is how it works. 05:43.000 --> 05:44.000 I mean just to summarize. 05:44.000 --> 05:48.000 So, ingestion, then pruning happen, pruning process happens. 05:48.000 --> 05:52.000 And then similar documents are grouped together, forming a cluster. 05:52.000 --> 05:54.000 And every cluster gets a summary. 05:54.000 --> 05:57.000 Lock 1, block 2, search happens. 05:57.000 --> 06:00.000 When the search happens, we query the summary vector. 06:00.000 --> 06:03.000 If the score is too low, the cluster is skipped. 06:03.000 --> 06:08.000 If the score is big enough, large enough, then we examine the cluster. 06:08.000 --> 06:10.000 And the cluster is examined. 06:10.000 --> 06:14.000 The full vector is gotten from the forward index, and the final scoring is time. 06:14.000 --> 06:19.000 So, this is how we are able to do approximate sparse retrieval. 06:19.000 --> 06:24.000 And this is how we are also able to scale beyond 10 million ceiling that we have in the past. 06:24.000 --> 06:25.000 Yeah. 06:25.000 --> 06:26.000 So, give it a shot. 06:26.000 --> 06:27.000 Give it a try. 06:27.000 --> 06:28.000 Let us know how it goes. 06:28.000 --> 06:29.000 You know the drill. 06:29.000 --> 06:30.000 Yeah. 06:30.000 --> 06:37.000 So, we three GTE is the Sata Infran3 pre-trained model that I was talking about. 06:37.000 --> 06:40.000 The most popular, most downloaded hugging case model. 06:40.000 --> 06:41.000 Yeah. 06:41.000 --> 06:42.000 And like I said, right. 06:42.000 --> 06:46.000 So, the benchmark that we did over 1.3 billion dot set. 06:46.000 --> 06:48.000 The benchmark that we did. 06:48.000 --> 06:50.000 Look at the box here. 06:50.000 --> 06:55.000 The latency with the BM25 is baseline. 06:55.000 --> 06:59.000 We were able to get four times faster than BM25. 06:59.000 --> 07:04.000 And about 11 times faster than the traditional neural sparse. 07:04.000 --> 07:05.000 Yeah. 07:05.000 --> 07:10.000 And the line about tells you the recall, which is a slight drop in recall. 07:10.000 --> 07:17.000 Has a such architect, such practitioner for me, a slight drop in recall, with such a benchmark 07:17.000 --> 07:18.000 latency that I get. 07:18.000 --> 07:20.000 It's a tremendous trade-off for me. 07:20.000 --> 07:21.000 I'm not sure what you. 07:21.000 --> 07:22.000 Yeah. 07:22.000 --> 07:23.000 So, yeah. 07:23.000 --> 07:24.000 Please try. 07:24.000 --> 07:25.000 Give it a shot. 07:25.000 --> 07:26.000 Yeah. 07:26.000 --> 07:28.000 So, I will move on to a new topic. 07:28.000 --> 07:30.000 System generated such pipeline. 07:30.000 --> 07:34.000 But in order for me to explain this one, you need to know that such pipelines exist with open 07:34.000 --> 07:35.000 search. 07:35.000 --> 07:38.000 And such pipelines are very similar to the ingest pipelines. 07:38.000 --> 07:42.000 A bunch of processors for data transformation that you put together. 07:42.000 --> 07:44.000 But on the set site, right. 07:44.000 --> 07:48.000 The only difference is the processors of different types. 07:48.000 --> 07:52.000 It could be request or response processor or a face processor. 07:52.000 --> 07:58.000 To give you some tangible examples, the normalization processor for normalizing the score 07:58.000 --> 08:00.000 between lexical and cane and search. 08:00.000 --> 08:05.000 And they cross-encorer for re-ranking and the semantic highlighter. 08:05.000 --> 08:06.000 Yeah. 08:06.000 --> 08:10.000 So, these are different processors that you put together to create a search pipeline. 08:10.000 --> 08:15.000 But the challenge is not the challenge, but the problem is always that you have to create 08:15.000 --> 08:21.000 them in advance, define them in advance, register them in advance before you are 08:21.000 --> 08:23.000 able to use them. 08:23.000 --> 08:24.000 Right. 08:24.000 --> 08:27.000 So, as the name suggests, the system generated such pipeline. 08:27.000 --> 08:31.000 It attempts to eliminate the need of having to create them in advance. 08:31.000 --> 08:37.000 So, to give you an example, if you are implementing e-commerce search, let's say you want to 08:37.000 --> 08:39.000 diversify your search results. 08:39.000 --> 08:42.000 And you want to have MMR, maximal, marginal relevancy. 08:42.000 --> 08:45.000 Now, you don't have to create these search pipelines in advance. 08:45.000 --> 08:48.000 You simply have to mention it in a such class. 08:48.000 --> 08:51.000 And the search pipeline is automatically created for you on the fly. 08:51.000 --> 08:55.000 So, it removes the operational overhead that you have. 08:55.000 --> 09:00.000 But the only setting is it is not available for all the search processes out there yet. 09:00.000 --> 09:01.000 Yeah. 09:01.000 --> 09:04.000 So, you still have to create the search pipelines. 09:04.000 --> 09:09.000 And for anyone who is new, there is a steep learning curve because you need to use the 09:09.000 --> 09:11.000 APIs to create the search pipelines. 09:11.000 --> 09:12.000 Yeah. 09:12.000 --> 09:18.000 So, once again to reduce the learning curve to cut out the learning curve, we unfor those who like 09:18.000 --> 09:20.000 the click-click create approach. 09:20.000 --> 09:21.000 So, we have this. 09:21.000 --> 09:23.000 So, you have, as you see here, right. 09:23.000 --> 09:27.000 So, different templates for hybrid search, multi-modal search, rag. 09:27.000 --> 09:28.000 And also agent search. 09:28.000 --> 09:29.000 Yeah. 09:29.000 --> 09:32.000 So, you have different templates which you can use. 09:32.000 --> 09:37.000 And you can also use the custom search if you want to, you know, go beyond the basics that we have. 09:37.000 --> 09:39.000 And it would look something like this. 09:39.000 --> 09:41.000 This is a full-blown screenshot. 09:41.000 --> 09:42.000 But they saved me. 09:42.000 --> 09:43.000 So, it goes like this. 09:43.000 --> 09:45.000 So, you create the ingest pipeline. 09:45.000 --> 09:46.000 Right. 09:46.000 --> 09:48.000 You can add a bunch of processes into it. 09:48.000 --> 09:52.000 I have the ML inference processor that can convert text to embeddings. 09:52.000 --> 09:55.000 And then, of course, you can add more processes. 09:55.000 --> 09:56.000 This does an example. 09:56.000 --> 09:57.000 Yeah. 09:57.000 --> 10:01.000 And then, we have the search pipeline where I have the same ML inference processor. 10:01.000 --> 10:04.000 I'm working to the text to embeddings. 10:04.000 --> 10:05.000 Yeah. 10:05.000 --> 10:08.000 And this is how you define what the processor should be doing. 10:08.000 --> 10:12.000 So, I then, it allows me to pick what model I want. 10:12.000 --> 10:16.000 Which field should I look at to convert the text. 10:16.000 --> 10:19.000 And once I convert the text to embeddings, where should I put it to. 10:19.000 --> 10:24.000 So, basically, you can configure the entire processor with the UI. 10:24.000 --> 10:25.000 Yeah. 10:25.000 --> 10:26.000 Not just that. 10:26.000 --> 10:28.000 You can also test. 10:28.000 --> 10:32.000 Simulates the entire flow in here before you start deploying it. 10:32.000 --> 10:33.000 Yeah. 10:33.000 --> 10:36.000 So, it's a very nice utility, a developer tool. 10:36.000 --> 10:37.000 Yeah. 10:37.000 --> 10:39.000 But let's say you are at this point. 10:39.000 --> 10:41.000 You have created the search pipeline. 10:41.000 --> 10:42.000 Everything is good. 10:42.000 --> 10:44.000 You have deployed the next thing that you need to know is. 10:44.000 --> 10:46.000 It's the search pipeline working at all. 10:46.000 --> 10:47.000 Yeah. 10:47.000 --> 10:51.000 So, enter search relevancy workbench and ask some tools. 10:51.000 --> 10:55.000 I would say, especially for your relevancy engineers, your data scientists. 10:55.000 --> 10:56.000 Yeah. 10:56.000 --> 10:58.000 So, as it shows here, right. 10:58.000 --> 11:00.000 You have four different things that you can do. 11:00.000 --> 11:02.000 You can compare single query. 11:02.000 --> 11:04.000 You can like side-by-side compare. 11:04.000 --> 11:06.000 I bought the search results. 11:06.000 --> 11:10.000 You can also do other volume using query set. 11:10.000 --> 11:14.000 You can also do search evaluation, which is basically all the. 11:14.000 --> 11:16.000 Good search quality metrics. 11:16.000 --> 11:18.000 The NDCGs and the positions. 11:18.000 --> 11:19.000 Yeah. 11:19.000 --> 11:23.000 And the hybrid search optimizer, which is slightly different from the rest of it. 11:23.000 --> 11:30.000 So, if you have used hybrid search, regardless of the vector database, you let's say you combine 11:30.000 --> 11:33.000 Mexico search and can and search. 11:33.000 --> 11:34.000 Right. 11:34.000 --> 11:35.000 So, you need to. 11:35.000 --> 11:38.000 There's a lot of parameters, hyper parameters that you need to tune. 11:38.000 --> 11:41.000 So, the normalization technique is that this core or min max or 11:41.000 --> 11:42.000 equilibrium. 11:42.000 --> 11:44.000 And then you have the combination technique. 11:44.000 --> 11:49.000 Is it arithmetic mean or geometric mean or harmonic mean? 11:49.000 --> 11:50.000 Yeah. 11:50.000 --> 11:53.000 And then lastly, you also have the combination. 11:53.000 --> 11:55.000 Distribution, which is. 11:55.000 --> 11:56.000 Legs equal to cannon. 11:56.000 --> 11:59.000 Is it 0.1.9.2.8? 11:59.000 --> 12:00.000 Yeah. 12:00.000 --> 12:04.000 So, you, you then compute the combinations that you would arrive at. 12:04.000 --> 12:07.000 That's way too many combinations. 12:07.000 --> 12:11.000 Then how would you know, this combination works the best for you. 12:11.000 --> 12:13.000 This is what the hybrid search optimizer does. 12:13.000 --> 12:18.000 I will show it in the demo, which is what I'm going to do right now. 12:18.000 --> 12:22.000 So, this is the new open search dashboard. 12:22.000 --> 12:30.000 This is not just the new look and feel, but it also lets you add different open search clusters 12:30.000 --> 12:32.000 to the same open search UI. 12:32.000 --> 12:38.000 So, this we go past the one dashboard to one backend. 12:38.000 --> 12:42.000 Now you are able to connect to different backends, open search clusters. 12:42.000 --> 12:47.000 And you can also create work spaces, a classic multi-tenancy use case, right? 12:47.000 --> 12:49.000 And then you create a workspace. 12:49.000 --> 12:53.000 It could also be of different types of observability or security analytics search 12:53.000 --> 12:54.000 depending upon it. 12:54.000 --> 12:56.000 The menu changes. 12:56.000 --> 13:00.000 And of course, if you want all the menus in the world like me, 13:00.000 --> 13:03.000 I go to the analytics. 13:03.000 --> 13:06.000 I get all the menus in the world, exactly. 13:06.000 --> 13:11.000 So, first thing I'm going to show you is I'm going to show you the agent search. 13:11.000 --> 13:14.000 And I'm going to compare that with the lexical search. 13:14.000 --> 13:19.000 So, for this, I go to the search relevancy work pens, single query comparison. 13:19.000 --> 13:20.000 Yeah? 13:20.000 --> 13:21.000 I have to type the query. 13:21.000 --> 13:23.000 Of course, I'm not going to do that now. 13:23.000 --> 13:25.000 I have it in here. 13:25.000 --> 13:30.000 And what you see here is I'm going to compare agent search with the multi-metge lexical search 13:30.000 --> 13:32.000 that you probably are used to. 13:32.000 --> 13:34.000 I'm going to use a local cluster for both. 13:34.000 --> 13:37.000 I'm going to use the e-commerce index for both. 13:37.000 --> 13:41.000 But of course, on this side, I'm going to use the agentic pipeline plot. 13:41.000 --> 13:43.000 This is a search pipeline that I created. 13:43.000 --> 13:44.000 Yeah? 13:44.000 --> 13:47.000 And I'm going to search for, let's say, Nike. 13:47.000 --> 13:48.000 Yeah? 13:48.000 --> 13:50.000 I'm going to hit a search. 13:50.000 --> 13:52.000 I have a very nice comparison here. 13:52.000 --> 13:55.000 There are three overlaps, seven unique results on each side. 13:55.000 --> 13:57.000 I'm going to change this to title. 13:57.000 --> 14:00.000 Make it a little bit bigger. 14:00.000 --> 14:03.000 So, the results are quite nice on both sides. 14:03.000 --> 14:05.000 But, of course, nobody is going to search for Nike. 14:05.000 --> 14:07.000 Nobody is going to search for jeans, right? 14:07.000 --> 14:09.000 So, they're going to be a bit more explicit, right? 14:09.000 --> 14:11.000 And I want to make a typo here, right? 14:11.000 --> 14:13.000 Nike black. 14:13.000 --> 14:20.000 Okay. 14:20.000 --> 14:21.000 Nike black shoes. 14:21.000 --> 14:23.000 Oh, one more unintentional typo. 14:23.000 --> 14:26.000 But, yeah, agentic search is able to handle it, right? 14:26.000 --> 14:28.000 So, if you now look at the results, right? 14:28.000 --> 14:29.000 This is no longer Nike. 14:29.000 --> 14:30.000 I mean, it is still shoes. 14:30.000 --> 14:31.000 Yeah? 14:31.000 --> 14:32.000 But, it is no longer Nike. 14:32.000 --> 14:36.000 On here, I have much more as a end user. 14:36.000 --> 14:40.000 I find this one much more relevant compared to the other one. 14:40.000 --> 14:41.000 Yeah? 14:41.000 --> 14:43.000 Now, let's continue the journey. 14:43.000 --> 14:49.000 Let's say under $50. 14:49.000 --> 14:50.000 Yeah? 14:50.000 --> 14:55.000 So, now, I no longer get shoes. 14:55.000 --> 14:56.000 I get cookies. 14:56.000 --> 14:57.000 I get whatnot. 14:57.000 --> 14:58.000 Yeah? 14:58.000 --> 15:01.000 Once again, on this side, I have much more relevant results. 15:01.000 --> 15:03.000 Of course, I show socks here. 15:03.000 --> 15:05.000 Also, you know why that is happening. 15:05.000 --> 15:08.000 But, also, $50, you get what you pay for, right? 15:08.000 --> 15:10.000 With this inflation. 15:10.000 --> 15:13.000 So, I'll show you what happens under the hood. 15:13.000 --> 15:14.000 There's no magic. 15:14.000 --> 15:15.000 Yeah? 15:15.000 --> 15:19.000 So, for this, we go to the AI search flow page. 15:19.000 --> 15:21.000 Remember, this is the one I showed you. 15:21.000 --> 15:25.000 And this is the one I use agentic search for creating this workflow. 15:25.000 --> 15:27.000 This is the one that I created. 15:27.000 --> 15:29.000 And the agent I have. 15:29.000 --> 15:31.000 This is the agent that I created. 15:31.000 --> 15:34.000 So, let's run the search first before I start explaining it. 15:34.000 --> 15:36.000 And this is exactly the same search that I did. 15:36.000 --> 15:40.000 And this is exactly what is causing the magic, right? 15:40.000 --> 15:44.000 So, agentic search is able to break down the user query. 15:44.000 --> 15:45.000 Understand the user query. 15:45.000 --> 15:48.000 Break it down into, let's say, 90 is a brand. 15:48.000 --> 15:50.000 Shoes are category. 15:50.000 --> 15:52.000 Black is a color. 15:52.000 --> 15:53.000 And price is a range. 15:53.000 --> 15:54.000 Yeah? 15:54.000 --> 15:55.000 And this is how. 15:55.000 --> 15:56.000 And this is the lexical search. 15:56.000 --> 15:57.000 There is no gain. 15:57.000 --> 15:58.000 And there is no sparsist. 15:58.000 --> 15:59.000 There's nothing. 15:59.000 --> 16:00.000 There's just lexical search. 16:00.000 --> 16:03.000 Just a powerful rewriting of the query, right? 16:03.000 --> 16:04.000 I get the raw results. 16:04.000 --> 16:07.000 The exact same results just to make sure that I'm not cheating. 16:07.000 --> 16:08.000 Yeah? 16:08.000 --> 16:11.000 So, now we go to the raw response, right? 16:11.000 --> 16:17.000 And we see in the category for the first result, right? 16:17.000 --> 16:21.000 So, clothing shoes, jewelry and athletic socks. 16:21.000 --> 16:22.000 No one there. 16:22.000 --> 16:24.000 This is actually showing a socks showing up in the shoe. 16:24.000 --> 16:26.000 Because there's a lot of cataloging problem. 16:26.000 --> 16:29.000 Not an agentic search problem, yeah? 16:29.000 --> 16:31.000 You also get the rewritten query here. 16:31.000 --> 16:33.000 But who's going to rewrite it? 16:33.000 --> 16:35.000 Which is what is happening on this side. 16:35.000 --> 16:37.000 So, we have this query planning tool. 16:37.000 --> 16:39.000 This is the tool for this agent. 16:39.000 --> 16:41.000 And as it says here, right? 16:41.000 --> 16:45.000 It is basically converting the natural language into open search DSL query. 16:45.000 --> 16:46.000 By using an LL. 16:46.000 --> 16:49.000 This is the LLM that I'm using. 16:49.000 --> 16:50.000 Yeah? 16:50.000 --> 16:52.000 Of course, I have another agent. 16:52.000 --> 16:54.000 Because this is open-source platform. 16:54.000 --> 16:59.000 So, we also have a agent that is actually running in Olamma 16:59.000 --> 17:03.000 on my tiny EC2 instance CPU, yeah? 17:03.000 --> 17:05.000 A 7 billion parameter. 17:05.000 --> 17:09.000 I'm saying all these things because it's super slow. 17:09.000 --> 17:10.000 Yeah? 17:10.000 --> 17:15.000 But just to prove a point, this is going to give me the exact query. 17:15.000 --> 17:17.000 If it doesn't fail me. 17:17.000 --> 17:21.000 But the bottom line is you are able to, you know? 17:21.000 --> 17:25.000 You don't need a super special cloud 4.5 model. 17:25.000 --> 17:30.000 You are able to do with a model that is just under 5 gigabytes. 17:30.000 --> 17:31.000 Yeah? 17:31.000 --> 17:34.000 So, you get the same exact query, same exact result. 17:34.000 --> 17:35.000 Yeah? 17:35.000 --> 17:39.000 So, and of course, one last thing I want to say is you can also, 17:39.000 --> 17:42.000 if you want to be a bit more deterministic. 17:42.000 --> 17:43.000 Yeah? 17:43.000 --> 17:45.000 Do not want to LLM to hallucinate. 17:45.000 --> 17:47.000 You can also use such templates. 17:47.000 --> 17:48.000 Yeah? 17:48.000 --> 17:50.000 You can take a bit more control. 17:50.000 --> 17:54.000 I think I'm taking more time than I may be running a bit late. 17:54.000 --> 17:57.000 But I also want to show you these things that I was talking about earlier. 17:57.000 --> 17:58.000 Right? 17:58.000 --> 18:00.000 So, we did a single query comparison. 18:00.000 --> 18:05.000 There is also a possibility to do a volume test with query set comparison. 18:05.000 --> 18:06.000 Right? 18:06.000 --> 18:07.000 Which is what I did here. 18:07.000 --> 18:08.000 Right? 18:08.000 --> 18:12.000 With 150 queries, you get all the ranking similarity metrics here. 18:12.000 --> 18:15.000 There is obviously a bug in here, which is fixed in the next release, 18:15.000 --> 18:17.000 which is what Jordan is going to talk about. 18:17.000 --> 18:20.000 But you basically can, you know, what you did at a single query. 18:20.000 --> 18:21.000 Right? 18:21.000 --> 18:23.000 Can be seen at a volume here. 18:23.000 --> 18:24.000 Right? 18:24.000 --> 18:29.000 But I like to actually show this one, the such event, right? 18:29.000 --> 18:34.000 Because for this one, if you go back, such evaluation, right? 18:34.000 --> 18:38.000 This is also volume as I'm going to pass a query here. 18:38.000 --> 18:40.000 This is just a field at query, right? 18:40.000 --> 18:44.000 But what I also am passing here is the judgment list. 18:44.000 --> 18:50.000 Because for such quality metrics for NDCG and for precision, you need judgment list. 18:50.000 --> 18:54.000 So, open search allows you to add either implicit or explicit judgments. 18:54.000 --> 18:57.000 Explicit, you just have to find it somehow. 18:57.000 --> 19:01.000 Implicit, you can use a project called user behavior inside, 19:01.000 --> 19:03.000 with an open search. 19:03.000 --> 19:07.000 It allows you to collect click metrics. 19:07.000 --> 19:11.000 Imagine Google Analytics, but such specific. 19:11.000 --> 19:13.000 So, take a look at the project. 19:13.000 --> 19:18.000 It's quite interesting, at least for me, as a such engineer. 19:18.000 --> 19:20.000 But go, yeah. 19:20.000 --> 19:24.000 The most nice thing is for your data scientists, 19:24.000 --> 19:30.000 to be able to look at it at a very thick step back, 19:30.000 --> 19:32.000 have a holistic picture. 19:32.000 --> 19:36.000 We paint a nice little dashboard for it. 19:36.000 --> 19:39.000 I'm not going to go over what is here, 19:39.000 --> 19:46.000 but just to make a point that you can install the dashboards, 19:46.000 --> 19:49.000 and you get this nice stuff. 19:49.000 --> 19:51.000 And lastly, the hyper search optimizer, 19:51.000 --> 19:54.000 I told you the different combinations for the hyper parameters, right? 19:54.000 --> 19:56.000 So, when you actually run it, 19:56.000 --> 19:59.000 you can actually see, T-Towels is a such query. 19:59.000 --> 20:02.000 It ran for 0.8.2, 20:02.000 --> 20:07.000 including an harmonic mean, 0.5.5, 20:07.000 --> 20:09.000 including a geometric mean, so on and so forth. 20:09.000 --> 20:10.000 You get the idea, right? 20:10.000 --> 20:12.000 So, it ran for every single query, 20:12.000 --> 20:15.000 every single combination of that, right? 20:15.000 --> 20:18.000 And then it tells me, let's go to the dashboards. 20:18.000 --> 20:21.000 Once again, it tells me, 20:21.000 --> 20:25.000 which combination is the best combination, 20:25.000 --> 20:28.000 best hyper parameter combination for me? 20:28.000 --> 20:32.000 For my data set, with the judgment list that I am giving, 20:32.000 --> 20:34.000 which is basically the arithmetic mean, 20:34.000 --> 20:37.000 equilibrium 0.1 point, whatever, I'm not able to see it. 20:37.000 --> 20:42.000 So, this is the power of this such relevancy workbench, 20:42.000 --> 20:48.000 try it out, and I think I will have to change the screen. 20:48.000 --> 20:52.000 So, thank you so much. 20:52.000 --> 20:54.000 I ask for a given round of applause. 20:54.000 --> 20:56.000 Come on. Here's the great job. 20:56.000 --> 20:59.000 And the demo goes with with us. 20:59.000 --> 21:02.000 So, we saw, we'll show you the QR code. 21:02.000 --> 21:06.000 So, you can actually run the demo yourself using the dev tools, 21:06.000 --> 21:10.000 API, or of course, through the open-setch dashboards UI, 21:10.000 --> 21:12.000 if it's more comfortable for you. 21:12.000 --> 21:16.000 So, you can do this, and it doesn't connect to the screen, 21:16.000 --> 21:18.000 external screen. 21:18.000 --> 21:24.000 And next up, we will show us some more search highlights in V3. 21:24.000 --> 21:25.000 Go ahead. 21:25.000 --> 21:26.000 Yeah. Thanks, Jordan. 21:26.000 --> 21:29.000 So, I think we are running out of time, 21:29.000 --> 21:31.000 but I want to tell you this one, right? 21:31.000 --> 21:34.000 So, Jordan was talking about the PPL, the pipe processing language. 21:34.000 --> 21:39.000 So, this is more or less a kind of a new programming, 21:39.000 --> 21:41.000 not programming, but query language. 21:41.000 --> 21:45.000 So, what we have is also a natural language ability 21:45.000 --> 21:47.000 to convert the natural language into PPL query. 21:47.000 --> 21:51.000 Yeah. So, as you see here in this nice little animation, right? 21:51.000 --> 21:54.000 So, you don't have to, you know, for using it, 21:54.000 --> 21:57.000 you don't have to know the query, get go, right? 21:57.000 --> 21:59.000 You can learn it along the way. 21:59.000 --> 22:02.000 But for more advanced users, oh, there's also nice little dashboards 22:02.000 --> 22:06.000 that is automatically created when you use the natural language to PPL. 22:07.000 --> 22:10.000 But for most more advanced users, yeah, 22:10.000 --> 22:12.000 you can also create the complicated queries. 22:12.000 --> 22:15.000 I will leave that there. I'm not going to say what kind of complicated queries. 22:15.000 --> 22:19.000 But there are also new commands that you can use, 22:19.000 --> 22:21.000 like Ragex. 22:21.000 --> 22:23.000 You can also use a Grok pattern. 22:23.000 --> 22:28.000 You can also use ML for camis and randomback forests. 22:28.000 --> 22:31.000 There's a lot of new commands that are possible with PPL. 22:31.000 --> 22:35.000 But the most striking thing for me is the ability to join. 22:35.000 --> 22:39.000 So, join is a new command that is released in PPL. 22:39.000 --> 22:42.000 Something that is long standing often asked. 22:42.000 --> 22:47.000 Now, you are able to, you know, join multiple indices in one, 22:47.000 --> 22:49.000 one one one place, yeah. 22:49.000 --> 22:51.000 Oh, sub query is also possible. 22:51.000 --> 22:55.000 I want to go with the rest of the stuff here, 22:55.000 --> 22:58.000 but I'm going to pass it over to. 22:58.000 --> 22:59.000 Thank you. 22:59.000 --> 23:01.000 So, I want to talk a bit about performance, 23:01.000 --> 23:04.000 because, you know, we're all about performance 23:04.000 --> 23:08.000 as being a point of emphasis for the open search community since its origins. 23:08.000 --> 23:12.000 A lot of users actually choose open search for its ability to ingest 23:12.000 --> 23:16.000 and manage and explore massive amounts of data, 23:16.000 --> 23:18.000 but with massive amounts of data comes obviously, 23:18.000 --> 23:20.000 a premium of the performance. 23:20.000 --> 23:24.000 So, we strive to build the most performant search engine out there, 23:24.000 --> 23:26.000 and which is obvious since version one. 23:26.000 --> 23:29.000 And this is why I brought to this plot from one. 23:29.000 --> 23:30.000 One.x to three. 23:30.000 --> 23:31.000 So, three.x, you can see. 23:31.000 --> 23:37.000 And actually, the gains have been so big that we build a logarithmic y-axis. 23:37.000 --> 23:39.000 So, when you see the graph going down, 23:39.000 --> 23:43.000 it's actually means that it's been decreasing exponentially. 23:43.000 --> 23:48.000 And, in fact, one that is across average across all query types, 23:48.000 --> 23:51.000 which is the big five workload that you probably know the benchmark, 23:51.000 --> 23:53.000 that's the dotted line in the middle, 23:53.000 --> 23:58.000 is improved 10 times from one.3 to three.x. 23:59.000 --> 24:01.000 So, pretty nice things to see. 24:01.000 --> 24:04.000 And by the way, it's not just about the query latency. 24:04.000 --> 24:09.000 We also see significant performance in the ingestion, 24:09.000 --> 24:13.000 in the storage, in throughput, in the transport. 24:13.000 --> 24:17.000 The dozens of new features in version three that contribute 24:17.000 --> 24:19.000 to these performance booths. 24:19.000 --> 24:21.000 So, we're running out of time. 24:21.000 --> 24:23.000 So, I won't go over all of these, 24:23.000 --> 24:25.000 but just to give you some highlights. 24:25.000 --> 24:26.000 First, the GPU acceleration. 24:26.000 --> 24:30.000 We only know here in the search dev room that vector operations 24:30.000 --> 24:35.000 are heavy, especially the distance calculations and the K&N. 24:35.000 --> 24:38.000 But, on the other hand, they lend themselves for parallel computation 24:38.000 --> 24:39.000 with GPUs. 24:39.000 --> 24:42.000 So, as part of the project, we've introduced 24:42.000 --> 24:45.000 a ability to leverage Nvidia architecture, 24:45.000 --> 24:48.000 the CVS Cagra algorithm, 24:48.000 --> 24:50.000 which shows pretty impressive results. 24:50.000 --> 24:55.000 You can see 9.3x improvement there on the indexing speed. 24:55.000 --> 24:57.000 Cost by nearly four times. 24:57.000 --> 24:59.000 That's on the GPU acceleration. 24:59.000 --> 25:01.000 Another cool thing is, 25:01.000 --> 25:03.000 we've seen over-ficed, 25:03.000 --> 25:05.000 because we know that we've seen, 25:05.000 --> 25:07.000 many of us know we've seen, 25:07.000 --> 25:09.000 it's memory-efficient, but doesn't scale very well. 25:09.000 --> 25:11.000 And on the other hand, 25:11.000 --> 25:13.000 scales very nicely, but then again, 25:13.000 --> 25:15.000 you need to load the entire 25:15.000 --> 25:17.000 Agenus W graph into memory. 25:17.000 --> 25:19.000 So, we came up with this clever thing of 25:19.000 --> 25:20.000 Lucine over-ficed. 25:20.000 --> 25:22.000 It gives the best of both worlds. 25:22.000 --> 25:24.000 Try it out, pretty new, 25:24.000 --> 25:25.000 learned for the CV. 25:25.000 --> 25:28.000 The initial test show impressive results. 25:28.000 --> 25:31.000 GRPC, with Protobuff, 25:31.000 --> 25:33.000 again, a very good way to, 25:33.000 --> 25:36.000 for faster, more efficient data transport 25:36.000 --> 25:40.000 and data processing compared to the JSON and rest. 25:40.000 --> 25:42.000 And especially with the 25:42.000 --> 25:44.000 significantly heavier operations, 25:44.000 --> 25:46.000 like bulk indexing, K&N, 25:46.000 --> 25:48.000 search and all of these. 25:48.000 --> 25:50.000 And also, it enables 25:50.000 --> 25:54.000 use of platforms such as Apache arrow and arrow fly. 25:54.000 --> 25:56.000 So, arrow, I don't know, 25:56.000 --> 25:58.000 the great talk yesterday for those 25:58.000 --> 25:59.000 attending databases there, 25:59.000 --> 26:00.000 but essentially, 26:00.000 --> 26:02.000 it eliminates the serialization and 26:02.000 --> 26:04.000 serialization overhead. 26:04.000 --> 26:06.000 And also, it allows for streaming architecture 26:06.000 --> 26:08.000 with processing aggregations, 26:08.000 --> 26:10.000 because it can allow to deliver 26:10.000 --> 26:13.000 partial results to boost complex aggregations. 26:13.000 --> 26:15.000 I took mentioned briefly before about 26:15.000 --> 26:17.000 the reader write a separation, 26:17.000 --> 26:19.000 so that you can separate the search and the indexing 26:19.000 --> 26:22.000 workloads to avoid the congestion and get 26:22.000 --> 26:25.000 more predictable performance. 26:25.000 --> 26:26.000 And many more, again, 26:26.000 --> 26:28.000 I won't have time to go over all of these. 26:28.000 --> 26:30.000 So, I want in the, 26:30.000 --> 26:33.000 to say a word about what's coming next. 26:33.000 --> 26:35.000 The opposite project builds on an eight week 26:35.000 --> 26:37.000 release train, so there's always 26:37.000 --> 26:40.000 good stuff coming around the corner. 26:40.000 --> 26:41.000 And in the coming month, 26:41.000 --> 26:43.000 we're going to release 3.5. 26:43.000 --> 26:45.000 And I wanted to just give you a taste of 26:45.000 --> 26:47.000 some of the innovations in search and 26:47.000 --> 26:48.000 observability. 26:48.000 --> 26:50.000 It is the ones that I particularly like. 26:50.000 --> 26:52.000 So, you know, 26:52.000 --> 26:53.000 but just 3.3, 26:53.000 --> 26:54.000 before last release, 26:54.000 --> 26:56.000 we've added skip lists for 26:56.000 --> 26:57.000 date histograms, 26:57.000 --> 26:58.000 for those who don't know. 26:58.000 --> 27:00.000 So, with 3.5, 27:00.000 --> 27:02.000 we're going to add skip list support for 27:02.000 --> 27:03.000 aggregations. 27:03.000 --> 27:05.000 So, for analytics use cases, 27:05.000 --> 27:07.000 this offers alternative for the 27:07.000 --> 27:09.000 multi-range traversal, 27:09.000 --> 27:11.000 that's much more efficient for 27:11.000 --> 27:13.000 certain frequently used query types. 27:13.000 --> 27:15.000 For example, if you have a query that 27:16.000 --> 27:18.000 filters on one field and then 27:18.000 --> 27:20.000 searches on another, 27:20.000 --> 27:21.000 things like that, 27:21.000 --> 27:24.000 we really see at least two times faster 27:24.000 --> 27:25.000 and so on. 27:25.000 --> 27:27.000 Another one is that the next release 27:27.000 --> 27:29.000 will introduce APM, 27:29.000 --> 27:30.000 a support application, 27:30.000 --> 27:32.000 performance monitoring with 27:32.000 --> 27:33.000 support for 27:33.000 --> 27:34.000 a permit years, 27:34.000 --> 27:35.000 for a metric data, 27:35.000 --> 27:37.000 open telemetry, 27:37.000 --> 27:39.000 as the instrumentation standard, 27:39.000 --> 27:41.000 so APM support really cool 27:41.000 --> 27:43.000 think coming next. 27:44.000 --> 27:45.000 And last but not least, 27:45.000 --> 27:46.000 it 3.5, 27:46.000 --> 27:47.000 we're going to introduce 27:47.000 --> 27:49.000 agent observability for 27:49.000 --> 27:50.000 root cause analysis. 27:50.000 --> 27:51.000 So, that's pretty cool, 27:51.000 --> 27:53.000 not having your agents as 27:53.000 --> 27:54.000 black boxes, 27:54.000 --> 27:55.000 we'll be able to analyze the 27:55.000 --> 27:56.000 performance, 27:56.000 --> 27:58.000 and refine our prompts, 27:58.000 --> 27:59.000 based on that. 27:59.000 --> 28:00.000 So, it's going to be released 28:00.000 --> 28:01.000 as experimentally, 28:01.000 --> 28:03.000 we want folks to start getting their hands on 28:03.000 --> 28:04.000 and give us feedback on the 28:04.000 --> 28:05.000 stand. 28:05.000 --> 28:06.000 It's all new to all of us, 28:06.000 --> 28:07.000 of course. 28:07.000 --> 28:08.000 So, to get a sense of this is 28:08.000 --> 28:10.000 the right direction as the community. 28:10.000 --> 28:11.000 And the old 3.5, 28:11.000 --> 28:12.000 there's a lot more, 28:12.000 --> 28:14.000 to talk about decoupling from 28:14.000 --> 28:16.000 Lucene and opening up for 28:16.000 --> 28:17.000 composable, 28:17.000 --> 28:19.000 pluggable architecture to introduce 28:19.000 --> 28:20.000 more engines, 28:20.000 --> 28:21.000 and more format, 28:21.000 --> 28:22.000 for example, 28:22.000 --> 28:23.000 per k, 28:23.000 --> 28:24.000 as a support, 28:24.000 --> 28:25.000 and much more. 28:25.000 --> 28:27.000 So, do stay tuned, 28:27.000 --> 28:28.000 and where is that, 28:28.000 --> 28:29.000 where it is, 28:29.000 --> 28:30.000 of course, 28:30.000 --> 28:31.000 the GitHub repo, 28:31.000 --> 28:32.000 and GitHub org with 28:32.000 --> 28:33.000 140 repos, 28:33.000 --> 28:34.000 actually, 28:34.000 --> 28:36.000 a very active Slack channel, 28:36.000 --> 28:37.000 instance, 28:37.000 --> 28:38.000 with thousands of participants, 28:38.000 --> 28:40.000 with our forums, 28:40.000 --> 28:41.000 and there's also the tags, 28:41.000 --> 28:42.000 technical advisory groups. 28:42.000 --> 28:44.000 I'm honored to be a lead for 28:44.000 --> 28:45.000 the observability tag. 28:45.000 --> 28:47.000 We have a monthly calls open for 28:47.000 --> 28:48.000 everyone, 28:48.000 --> 28:49.000 and Slack channel for that. 28:49.000 --> 28:50.000 So, if you're into that, 28:50.000 --> 28:51.000 do join. 28:51.000 --> 28:52.000 We have us, 28:52.000 --> 28:53.000 groups around the world, 28:53.000 --> 28:54.000 around 19 countries, 28:54.000 --> 28:55.000 I think, around the world. 28:55.000 --> 28:56.000 And also, 28:56.000 --> 28:58.000 the open source content events 28:58.000 --> 28:59.000 around the world. 28:59.000 --> 29:00.000 I just came back last month 29:00.000 --> 29:01.000 from OpenSedgeConjapan. 29:01.000 --> 29:03.000 That's the bonsai tree here. 29:03.000 --> 29:04.000 And we're going to have 29:04.000 --> 29:05.000 openSedgeCon Europe, 29:05.000 --> 29:06.000 assuming most of you are 29:06.000 --> 29:07.000 based in Europe, 29:07.000 --> 29:08.000 in April, 29:08.000 --> 29:09.000 in Prague. 29:09.000 --> 29:10.000 So, you're all invited to join, 29:10.000 --> 29:11.000 by the way. 29:11.000 --> 29:13.000 It has a special track for a patchy 29:13.000 --> 29:14.000 loose scene and search. 29:14.000 --> 29:15.000 So, even if you're not 29:15.000 --> 29:16.000 specifically with OpenSedge, 29:16.000 --> 29:17.000 but the broader search 29:17.000 --> 29:18.000 community, 29:18.000 --> 29:19.000 we welcome your participation, 29:19.000 --> 29:21.000 also CFP for the following 29:21.000 --> 29:22.000 ones, 29:22.000 --> 29:23.000 check it out. 29:23.000 --> 29:25.000 And check out the QR code, 29:25.000 --> 29:26.000 all the links are there 29:26.000 --> 29:27.000 in one page, 29:27.000 --> 29:28.000 including some other stuff, 29:28.000 --> 29:29.000 like roadmap, 29:29.000 --> 29:31.000 like the playgrounds 29:31.000 --> 29:32.000 that you can play online, 29:32.000 --> 29:33.000 and so on. 29:33.000 --> 29:34.000 And of course, 29:34.000 --> 29:35.000 Ash Wattenme, 29:35.000 --> 29:36.000 are here, 29:36.000 --> 29:37.000 happy to answer any questions 29:37.000 --> 29:38.000 after the talk. 29:38.000 --> 29:39.000 So, after the conference, 29:39.000 --> 29:40.000 so we put the QR code, 29:40.000 --> 29:41.000 so you can contact us 29:41.000 --> 29:42.000 also online, 29:42.000 --> 29:43.000 if we don't get to chat here. 29:43.000 --> 29:44.000 So, feel free to contact 29:44.000 --> 29:45.000 we'd love to hear your 29:45.000 --> 29:46.000 question, 29:46.000 --> 29:47.000 you feedback about this talk, 29:47.000 --> 29:48.000 and anything in between. 29:48.000 --> 29:49.000 I'm Dotaan, 29:49.000 --> 29:50.000 this is Ash Wattenme. 29:50.000 --> 29:51.000 Thank you very much for listening. 29:51.000 --> 29:52.000 Thank you. 29:52.000 --> 29:55.000 Do you have 30 seconds?