WEBVTT 00:00.000 --> 00:12.680 So, the next talk will be presented by Sid Rikul Mer, and it's going to be about end-to-end 00:12.680 --> 00:15.000 enterprise search with Datafariq. 00:15.000 --> 00:16.000 Yes, good. 00:16.000 --> 00:22.560 So, I have the immense privilege of having 25 minutes, and my laptop worked right away. 00:22.560 --> 00:29.960 So, when you have finished on time, so I'm Sid Rik and the President co-founder of 00:30.040 --> 00:34.760 France Labs, who has a software maker of Datafariq, and so today, we're going to talk 00:34.760 --> 00:39.560 about end-to-end enterprise search with Datafariq community edition, because we have two 00:39.560 --> 00:40.560 editions. 00:40.560 --> 00:46.960 So, Viagenda, we'll talk a bit about France Labs, organization, definition of enterprise 00:46.960 --> 00:50.960 search, because I'm not sure if we're all in line with what it means. 00:50.960 --> 00:56.960 Let me talk about Datafariq's architecture, the main functionality is, and some Gen AI, 00:57.960 --> 01:02.960 I was going to say that everybody mentions AI, but passable did not mention AI, so it's an exception. 01:02.960 --> 01:06.960 Maybe did not have time too, so you're ready? 01:06.960 --> 01:07.960 Yeah. 01:07.960 --> 01:08.960 Oh, yeah, wow. 01:08.960 --> 01:09.960 Okay, thank you. 01:09.960 --> 01:13.960 So, we were born 40, so who knew Datafariq before? 01:13.960 --> 01:15.960 Yeah, I think you're expecting that. 01:15.960 --> 01:16.960 That's good. 01:16.960 --> 01:18.960 So, we are specialising not being known. 01:18.960 --> 01:23.960 So, let's see, for 14 years of existence, we released Datafariq in 2015. 01:23.960 --> 01:28.960 November, we're basing these on the French Riviera, with a swimming pool. 01:28.960 --> 01:31.960 We speak in many conferences related to search. 01:31.960 --> 01:34.960 That's why maybe you don't know who you are, enterprise search. 01:34.960 --> 01:38.960 We are involved in open source businesses, so we are part of the French, 01:38.960 --> 01:44.960 Cluster, French Association of Open Source Businesses, and European Association of Open Source Businesses, 01:44.960 --> 01:47.960 respectively, CNL and Appell. 01:47.960 --> 01:53.960 Members of the search network and customers in weird industry sectors. 01:53.960 --> 01:58.960 So, the surprise, with sort of a make of Datafariq, so we provide expertise around it, 01:58.960 --> 02:00.960 and also an Apache solar. 02:00.960 --> 02:02.960 So, what isn't the price search? 02:02.960 --> 02:08.960 So, I could be pasted via, what I'm quoting, actually, the definition for Martin White, 02:08.960 --> 02:11.960 and he's book enterprise search, actually, that's been passed here. 02:11.960 --> 02:14.960 He was the only serious book about enterprise search, so it's. 02:14.960 --> 02:18.960 The enterprise search application enables employees to find all the information, 02:18.960 --> 02:22.960 but the company processes without the need to know where information is stored. 02:22.960 --> 02:26.960 And I did, oops, sorry, I didn't hit the microphone, but, 02:26.960 --> 02:29.960 particularly because most customers, we want security. 02:29.960 --> 02:34.960 By security, I mean, here, respecting the access control to documents. 02:34.960 --> 02:38.960 So, what does the typical enterprise search solution works? 02:38.960 --> 02:42.960 Well, it looks like a search solution, no surprise here. 02:42.960 --> 02:46.960 So, we'll detail a bit more what it does in Datafariq, but, yeah, it's a search bar. 02:46.960 --> 02:51.960 You search, you have a complete, you have facets to filter, and you have a results list. 02:51.960 --> 02:52.960 Okay, no surprise. 02:52.960 --> 02:56.960 Any addition, you have some monitoring, like statistics, analytics, 02:56.960 --> 03:00.960 dashboards to see who search is for what? 03:00.960 --> 03:03.960 And how does that usually work? 03:03.960 --> 03:06.960 Well, there are several components from left to right. 03:06.960 --> 03:09.960 You have a crawling face where you connect to the data sources, 03:09.960 --> 03:11.960 the documents, whoever they are stored. 03:11.960 --> 03:13.960 You retrieve them, you open them, you analyze what's in it, 03:13.960 --> 03:17.960 you take the metadata, you take the content, you put it in an index, 03:17.960 --> 03:19.960 search index, then you have the search components, 03:19.960 --> 03:23.960 we take a care of in real time analyzing the user's query, 03:23.960 --> 03:27.960 looking for the data in Vindex, and providing 03:27.960 --> 03:32.960 the results usually in a UI, and also a component for statistics. 03:32.960 --> 03:36.960 Okay, so now you are experts in enterprise search. 03:36.960 --> 03:38.960 Yeah. 03:38.960 --> 03:42.960 Okay, so what the data far is CE allows to do. 03:42.960 --> 03:44.960 So CE is the community edition. 03:44.960 --> 03:46.960 Because it's first them, I don't know. 03:46.960 --> 03:48.960 You choose for a Rocks at me. 03:48.960 --> 03:52.960 So it allows you to crawl multiple types of documents, 03:52.960 --> 03:54.960 and I have some slides explaining a bit more what it does, 03:54.960 --> 03:57.960 and this is thanks to Apache many full CF, 03:57.960 --> 04:01.960 which is two minor ledge, we only existing open source framework, 04:01.960 --> 04:03.960 doing this type of stuff. 04:03.960 --> 04:06.960 We use Apache Chica, we'll see that afterwards. 04:07.960 --> 04:11.960 Data Free CE allows you to do some search using VM25, 04:11.960 --> 04:16.960 which is SQL2TFIDF, and so we will do vector search. 04:16.960 --> 04:20.960 I will talk about it later, and while that we use Apache Solar, 04:20.960 --> 04:24.960 you know Apache Solar, we're in good, 04:24.960 --> 04:27.960 some nodding here, so that's good. 04:27.960 --> 04:30.960 So you are able to do some analysis, 04:30.960 --> 04:33.960 thanks to Dashboards, and you're able to manage your tool, 04:33.960 --> 04:36.960 which is difficult to use interfaces for administrators, 04:36.960 --> 04:40.960 from playing with the weights, doing some promo links. 04:40.960 --> 04:43.960 I don't have time for that, but entity detection, etc. 04:43.960 --> 04:47.960 You can use the admin UI of many full CF and Apache Solar. 04:47.960 --> 04:49.960 So what does it not allow you to do? 04:49.960 --> 04:52.960 So I put easily, because we don't form it anything. 04:52.960 --> 04:55.960 Everything's open, so you can find the time to do it, 04:55.960 --> 04:58.960 just that in order to survive, this is what I post up, 04:58.960 --> 05:01.960 but we put as extra intent to price edition. 05:01.960 --> 05:04.960 And security, and here I'm talking about access control, 05:04.960 --> 05:06.960 paramission document-level permissions, 05:06.960 --> 05:09.960 and to connect to a key directory, 05:09.960 --> 05:11.960 they want to adapt this type of stuff. 05:11.960 --> 05:13.960 Does not easily allow you to do big data search, 05:13.960 --> 05:16.960 in the sense that if you need a cluster of data for a system 05:16.960 --> 05:18.960 with distributed components, 05:18.960 --> 05:20.960 this is not done automatically, 05:20.960 --> 05:22.960 there is no answer, but recipient is kind of stuff. 05:22.960 --> 05:24.960 And it's not easy to do a backup, 05:24.960 --> 05:27.960 well, a nest, of course, if you are the Docker, 05:27.960 --> 05:30.960 and it's persistent, you can just save it, but then. 05:30.960 --> 05:32.960 And we don't prevent anything, okay. 05:32.960 --> 05:34.960 So I hope you feel better now. 05:34.960 --> 05:37.960 So what does it look like in terms of architecture? 05:37.960 --> 05:40.960 So it's one, it's on our weeky. 05:40.960 --> 05:43.960 And yeah, so that seems a gray box, 05:43.960 --> 05:46.960 that's for the monolithic version of Datafari. 05:46.960 --> 05:48.960 So in the gray box, it's exposing, 05:48.960 --> 05:51.960 it's using interface to users on the left, 05:51.960 --> 05:55.960 and we fetch the data, which on the right of the schema. 05:55.960 --> 05:59.960 So you can see we have some Apache, some Tomcat, 05:59.960 --> 06:03.960 some Zuki Perse, some Solos, some Casanoa, some Puzz Gray, 06:03.960 --> 06:05.960 LogStash, for historical reasons, 06:05.960 --> 06:09.960 Tika and I think, I've mentioned all of them. 06:09.960 --> 06:12.960 And it's very, it's a big soup of stuff to manage 06:12.960 --> 06:13.960 and to deploy. 06:13.960 --> 06:16.960 So it lots of fun when doing the deployments scripts, 06:16.960 --> 06:18.960 but we work it out. 06:18.960 --> 06:20.960 So the main components for that, 06:20.960 --> 06:22.960 we're using, so really the main ones 06:22.960 --> 06:25.960 would be Apache mini 4 CF, because that's the very framework 06:25.960 --> 06:27.960 for crawling the documents. 06:27.960 --> 06:30.960 And it's Apache Solos, for the search part, 06:30.960 --> 06:33.960 and it's Apache Tomcat, for owned Datafari web app. 06:33.960 --> 06:38.960 And we provide an admin UI to install managing use. 06:38.960 --> 06:41.960 If we dig a bit into many false CF, 06:41.960 --> 06:43.960 anybody new, many false CF people? 06:43.960 --> 06:44.960 Yeah, no. 06:44.960 --> 06:45.960 No, okay. 06:45.960 --> 06:48.960 Nobody cares about crawling, but it's an important step in, 06:48.960 --> 06:50.960 oops, let's keep it. 06:50.960 --> 06:52.960 So many false CF allows you to connect, 06:52.960 --> 06:54.960 to, I don't know, for instance, I put, 06:54.960 --> 06:56.960 yeah, file shares, I take the directory, 06:56.960 --> 06:58.960 I'm an LDAP web, XWiki. 06:58.960 --> 06:59.960 I mean, I think all of us, 06:59.960 --> 07:01.960 many presentations today were mentioned, 07:01.960 --> 07:03.960 this week is, I did my homework. 07:03.960 --> 07:05.960 We have a connect of YWiki. 07:05.960 --> 07:09.960 Yeah, you can have a pipeline of transformation 07:09.960 --> 07:11.960 connectors to transform the dates, 07:11.960 --> 07:12.960 have a choice of extracted. 07:12.960 --> 07:14.960 Maybe you want to put some regular expressions 07:14.960 --> 07:17.960 to find entities, maybe you want to 07:17.960 --> 07:19.960 print out some particular document types, 07:19.960 --> 07:21.960 or, yeah, I don't know, 07:21.960 --> 07:24.960 if things can be done in this pipeline, 07:24.960 --> 07:27.960 and then we push whatever has been analyzed 07:27.960 --> 07:28.960 into the solar index. 07:28.960 --> 07:29.960 So that's how many false CFs, 07:29.960 --> 07:31.960 so you can see many false CFs 07:31.960 --> 07:34.960 as an octopuses able to connect to many types. 07:34.960 --> 07:37.960 Many types of data management systems. 07:37.960 --> 07:39.960 It's also a month array. 07:39.960 --> 07:41.960 I didn't draw them, 07:41.960 --> 07:44.960 but they are coupé left, I think I hope. 07:44.960 --> 07:47.960 I think the guy is dead when it was years ago, 07:47.960 --> 07:48.960 so that's fine. 07:48.960 --> 07:50.960 So he can open me formats, 07:50.960 --> 07:53.960 to Apache Tika, so PDF, obviously, 07:53.960 --> 07:55.960 liberal fees, Microsoft Office, 07:55.960 --> 07:58.960 database tables, any other types. 07:58.960 --> 08:03.960 So it goes quick, but that's fine. 08:03.960 --> 08:06.960 So now you're expressing that patchymane false CF. 08:06.960 --> 08:09.960 Yeah, of course, now about solar. 08:09.960 --> 08:12.960 So well, solar is job is to take what a patchymane 08:12.960 --> 08:14.960 false CF sends it, 08:14.960 --> 08:17.960 and analyze it and create the solar documents 08:17.960 --> 08:18.960 that it puts in Vindex, 08:18.960 --> 08:20.960 and it's job is then at the search phase 08:20.960 --> 08:23.960 to answer very fast to the user's query, 08:23.960 --> 08:27.960 and to deliver relevant documents classified. 08:27.960 --> 08:29.960 It's similar to open search, 08:29.960 --> 08:32.960 if you just know open search with them. 08:32.960 --> 08:34.960 And for the analytics components, 08:34.960 --> 08:36.960 well, but now it has become pretty easy, 08:36.960 --> 08:37.960 what we do, because in the past, 08:37.960 --> 08:38.960 we were using L, 08:38.960 --> 08:39.960 and then we were fed up with L, 08:39.960 --> 08:41.960 and then we used Zeppelin, 08:41.960 --> 08:43.960 and Zeppelin was grass roots. 08:43.960 --> 08:45.960 So now we got rid of all that, 08:45.960 --> 08:47.960 and we take everything that's come from 08:47.960 --> 08:49.960 from the logs of Apache Manifold CF, 08:49.960 --> 08:50.960 of Apache solar. 08:50.960 --> 08:52.960 We still have log-stash, 08:52.960 --> 08:53.960 we need to get rid of it, 08:53.960 --> 08:56.960 and we put it in the dedicated search collection of solar, 08:56.960 --> 08:59.960 and this is where how we feed our home-made 08:59.960 --> 09:02.960 dash balls in pure HTML. 09:02.960 --> 09:04.960 Also, we stopped frameworks, 09:04.960 --> 09:06.960 like Zeppelin. 09:06.960 --> 09:09.960 So if we detail a bit, 09:09.960 --> 09:11.960 what you currently have with the data for UI, 09:11.960 --> 09:15.960 this is the screenshot of the version 2 of React UI. 09:16.960 --> 09:18.960 It's actually very similar to previous ones. 09:18.960 --> 09:20.960 So this is, and it cost us a lot to do, 09:20.960 --> 09:22.960 but we forgot to modify, 09:22.960 --> 09:24.960 we look and feel so that you have the feeling it's brand new. 09:24.960 --> 09:27.960 So it's brand new, a bit like the old one. 09:27.960 --> 09:30.960 So you have a search bar on top. 09:30.960 --> 09:31.960 When you start typing, 09:31.960 --> 09:33.960 you have your dropdown that appears, 09:33.960 --> 09:34.960 or you can also remove it, 09:34.960 --> 09:35.960 but that's fine. 09:35.960 --> 09:37.960 You have the suggestions for queries, 09:37.960 --> 09:38.960 or the standard, 09:38.960 --> 09:39.960 or the complete, 09:39.960 --> 09:40.960 you can do, 09:40.960 --> 09:41.960 we can have auto suggestions, 09:41.960 --> 09:43.960 which is proposing documents right away, 09:43.960 --> 09:45.960 to add some suggestions based on specific fields. 09:45.960 --> 09:46.960 So if you have product numbers, 09:46.960 --> 09:47.960 for instance, 09:47.960 --> 09:49.960 if it could propose you product numbers, 09:49.960 --> 09:50.960 as well, to complete. 09:50.960 --> 09:52.960 So that's why we drop down. 09:52.960 --> 09:54.960 Then you can see a bit on the left, 09:54.960 --> 09:55.960 the tabs, 09:55.960 --> 09:59.960 which allow you to filter based on a specific type of metadata that you want. 09:59.960 --> 10:01.960 You can also opt for facets. 10:01.960 --> 10:04.960 Any metadata can be a facet, 10:04.960 --> 10:05.960 basically. 10:05.960 --> 10:07.960 And you have different types of interactions with your facets. 10:07.960 --> 10:09.960 Can we have a checkboxes? 10:09.960 --> 10:11.960 Can we have the Windows, 10:11.960 --> 10:13.960 for example time window, or price window, 10:13.960 --> 10:14.960 whatever you want, 10:14.960 --> 10:15.960 or if you missed that, 10:15.960 --> 10:16.960 wait too much. 10:16.960 --> 10:17.960 Thank you. 10:17.960 --> 10:20.960 I hope you have lots of questions of the one. 10:20.960 --> 10:22.960 Otherwise, I will ask you questions. 10:22.960 --> 10:24.960 So this I covered. 10:24.960 --> 10:25.960 You had the search tools. 10:25.960 --> 10:27.960 I didn't put the screenshot about it, 10:27.960 --> 10:29.960 but you can export your search results. 10:29.960 --> 10:31.960 You can save your searches. 10:31.960 --> 10:33.960 You can have email alerts. 10:33.960 --> 10:35.960 So if you don't want every day to come to your corpus, 10:35.960 --> 10:37.960 and ask that data vary, 10:37.960 --> 10:39.960 did the new document appear. 10:39.960 --> 10:42.960 Then the data vary will run these queries for you, 10:42.960 --> 10:43.960 and send you an email saying, 10:43.960 --> 10:45.960 if today there is something for you. 10:45.960 --> 10:47.960 And you can do advanced search, 10:47.960 --> 10:50.960 graphically to combine search for specific fields 10:50.960 --> 10:52.960 with Boolean operators. 10:52.960 --> 10:53.960 Okay, let's clear. 10:53.960 --> 10:54.960 You understood using the face. 10:54.960 --> 10:56.960 There's specific challenge normally, 10:56.960 --> 10:57.960 but we never know. 10:57.960 --> 10:58.960 And you can notice, 10:58.960 --> 11:00.960 little box asks your content. 11:00.960 --> 11:01.960 It's a teaser, 11:01.960 --> 11:03.960 because we will go back to it a bit later. 11:03.960 --> 11:05.960 But you can see it's fair. 11:05.960 --> 11:07.960 Ask your content. 11:07.960 --> 11:11.960 I'm seeing this because now it's time to talk about AI. 11:11.960 --> 11:13.960 So historically we could connect 11:13.960 --> 11:15.960 at the indexing phase to install 11:15.960 --> 11:18.960 other field party systems to do a translation, 11:18.960 --> 11:20.960 or imagine analysis. 11:20.960 --> 11:23.960 In our case, most of the time 11:23.960 --> 11:25.960 it was to do optical character recognition. 11:25.960 --> 11:29.960 We've test your act for for image recognition. 11:29.960 --> 11:31.960 And now it's kind of changing, 11:31.960 --> 11:33.960 and it's more focused on using 11:34.960 --> 11:36.960 Gen AI to do this type of stuff. 11:36.960 --> 11:38.960 So it's actually easier for us. 11:38.960 --> 11:40.960 And so we can connect 11:40.960 --> 11:42.960 at indexing time and at search time. 11:42.960 --> 11:43.960 And at search time, 11:43.960 --> 11:46.960 it's more for the vector search and the rug. 11:46.960 --> 11:49.960 And I will explain a bit more about the rug afterwards. 11:49.960 --> 11:52.960 Visorinjectivity is actually so not co-founded, 11:52.960 --> 11:55.960 but co-funded by European Union. 11:55.960 --> 11:58.960 More precisely the NNGI search project, 11:58.960 --> 12:00.960 called Neural Data Faber project. 12:01.960 --> 12:03.960 View maze by Vologo. 12:03.960 --> 12:04.960 Which is a great logo. 12:04.960 --> 12:06.960 And the N2N prototype is planned 12:06.960 --> 12:08.960 for what we end of this quarter. 12:08.960 --> 12:10.960 So we think we need to hurry for that. 12:10.960 --> 12:13.960 Visorinject is done in collaboration 12:13.960 --> 12:16.960 with CIS, S-E-A-S-E, 12:16.960 --> 12:18.960 company in the UK. 12:18.960 --> 12:21.960 More precisely Alessandro Benedetti, 12:21.960 --> 12:23.960 who is a committer in Apache Solar. 12:23.960 --> 12:25.960 And actually yesterday evening 12:25.960 --> 12:29.960 he did the second pull request on Solar 12:29.960 --> 12:36.960 to add VMBedit's call of the embedding model in Solar. 12:36.960 --> 12:38.960 So the PR has been done, 12:38.960 --> 12:40.960 now it's waiting for reviews, 12:40.960 --> 12:42.960 which means that from now on in Solar 12:42.960 --> 12:44.960 independently from data ferry, 12:44.960 --> 12:45.960 you just need to push the document, 12:45.960 --> 12:46.960 actually like the parameter, 12:46.960 --> 12:48.960 say which model you want, 12:48.960 --> 12:51.960 and Solar will take care of doing the embedding. 12:51.960 --> 12:53.960 Until yesterday it was not the case. 12:53.960 --> 12:55.960 Until yesterday you had to do the embedding externally. 12:55.960 --> 12:58.960 And then Solar was able to do the vector calculations 12:58.960 --> 12:59.960 and vector management, 12:59.960 --> 13:01.960 but not the embedding itself. 13:01.960 --> 13:04.960 So it's quite fresh actually. 13:04.960 --> 13:06.960 So Rag, what does it look like? 13:06.960 --> 13:08.960 So we're going back to videos into phase. 13:08.960 --> 13:11.960 And so the little box I should do on the lower rights, 13:11.960 --> 13:14.960 if you click on it, it's a chatbot. 13:14.960 --> 13:16.960 I'm sure it's not really a surprise for it, 13:16.960 --> 13:19.960 but what does it allow you to do? 13:19.960 --> 13:21.960 You can ask your content. 13:21.960 --> 13:25.960 So the first mechanism is the typical Rag 13:26.960 --> 13:28.960 recovery management generation. 13:28.960 --> 13:31.960 Everybody's familiar with Hague. 13:31.960 --> 13:33.960 Yeah, I have some handshake. 13:33.960 --> 13:36.960 I will take it as a yes for everybody. 13:36.960 --> 13:41.960 So we go as a Rag for all of the content. 13:41.960 --> 13:44.960 But you can also ask questions for specific documents. 13:44.960 --> 13:45.960 So in the search results, 13:45.960 --> 13:48.960 you can notice that there is a little ask with AI. 13:48.960 --> 13:50.960 But then if you pick on it, 13:50.960 --> 13:53.960 the chatbot changes and allows you to ask questions 13:53.960 --> 13:55.960 only for this particular document. 13:55.960 --> 13:57.960 Because I mean, we have customers. 13:57.960 --> 14:00.960 We have documents which are 6,000 pages long. 14:00.960 --> 14:02.960 Which is not a good idea actually. 14:02.960 --> 14:04.960 I think they should do several books. 14:04.960 --> 14:05.960 But okay. 14:05.960 --> 14:07.960 And at least with that, 14:07.960 --> 14:09.960 you can ask a question only for that. 14:09.960 --> 14:11.960 And we have ideals for quick actions. 14:11.960 --> 14:14.960 So it's prepared actually prompt for instance 14:14.960 --> 14:16.960 to do the summarization of a specific document. 14:16.960 --> 14:18.960 And then you can decide. 14:18.960 --> 14:21.960 You can create your own quick actions. 14:22.960 --> 14:25.960 This is meant for for his purpose. 14:25.960 --> 14:26.960 So what is Rags? 14:26.960 --> 14:27.960 I knew you said you knew it, 14:27.960 --> 14:28.960 but just behind the scenes. 14:28.960 --> 14:29.960 So it's a standard. 14:29.960 --> 14:31.960 So user, I did the drawings. 14:31.960 --> 14:33.960 I hope you're amazed. 14:33.960 --> 14:34.960 Okay. 14:34.960 --> 14:35.960 Really proud of them, 14:35.960 --> 14:36.960 especially for brain. 14:36.960 --> 14:37.960 It's amazing. 14:37.960 --> 14:39.960 So we use it as a query. 14:39.960 --> 14:41.960 He's happy before doing the query. 14:41.960 --> 14:43.960 But so data for it as a vector search. 14:43.960 --> 14:44.960 Thank you. 14:44.960 --> 14:46.960 So to the so-law, we request dimensions. 14:46.960 --> 14:48.960 We have a dedicated collection. 14:48.960 --> 14:50.960 Composive mini chunks. 14:51.960 --> 14:54.960 Of the vectors coming from the documents that I've been indexed. 14:54.960 --> 14:56.960 We classify them. 14:56.960 --> 14:59.960 We take the top end depending on the context of the capabilities. 14:59.960 --> 15:02.960 And we ask for GNAI. 15:02.960 --> 15:05.960 Please answer to the user's query. 15:05.960 --> 15:10.960 Using only these chunks as a context to the prompt. 15:10.960 --> 15:17.960 You can either ask an onsite or a cloud cloud, 15:17.960 --> 15:18.960 trade LLM. 15:18.960 --> 15:22.960 As long as it's compatible with OpenAI API interface. 15:22.960 --> 15:24.960 For instance, Mistral. 15:24.960 --> 15:25.960 Yeah. 15:25.960 --> 15:27.960 To promote some French innovation. 15:27.960 --> 15:30.960 I know he's just a gentleman one. 15:30.960 --> 15:32.960 I think it collapsed. 15:32.960 --> 15:34.960 No gentleman LLM. 15:34.960 --> 15:35.960 You don't know. 15:35.960 --> 15:36.960 There is no gentleman in the room. 15:36.960 --> 15:37.960 So that's fine. 15:37.960 --> 15:38.960 No. 15:38.960 --> 15:39.960 It's not good. 15:39.960 --> 15:40.960 Perfect. 15:40.960 --> 15:42.960 And then we have LLM generates a nice answer. 15:42.960 --> 15:45.960 And data for it takes it generates a user interface. 15:46.960 --> 15:51.960 The way it works, now that we have the Solovic top part. 15:51.960 --> 15:52.960 So why do we do rag? 15:52.960 --> 15:55.960 Well, why not proposing customers to train LLM. 15:55.960 --> 15:59.960 So there are these are the rationales for the enterprise edition. 15:59.960 --> 16:02.960 And this is the rational for the community edition. 16:02.960 --> 16:08.960 So the enterprise edition is because if you train LLM with the copies of the commands, 16:08.960 --> 16:11.960 LLM does not care about permission level systems. 16:11.960 --> 16:13.960 It will generate the statistics of words. 16:13.960 --> 16:16.960 And it will create an answer out of the complete copies. 16:16.960 --> 16:21.960 So you will disclose confidential information to potentially to an employee. 16:21.960 --> 16:23.960 So you can't, you can't do that. 16:23.960 --> 16:27.960 And we flag because we have this first step of doing the vector search. 16:27.960 --> 16:29.960 Based on the documents that Solov manages. 16:29.960 --> 16:33.960 We do filter out the documents that the users are not allowed to see. 16:33.960 --> 16:38.960 And this way, the answer generated is only based on what the user is allowed to see. 16:38.960 --> 16:42.960 But for the customer of the community edition, 16:42.960 --> 16:48.960 it's not easy to manage modifications in an LLM when you train it. 16:48.960 --> 16:50.960 Imagine you have train it with documents. 16:50.960 --> 16:52.960 And now the documents need to be modified. 16:52.960 --> 16:54.960 You cannot just say to be an M4, get this document. 16:54.960 --> 16:56.960 And here is a new version. 16:56.960 --> 16:57.960 Okay, there are many things to do. 16:57.960 --> 16:59.960 It was basically one being to retrain everything, 16:59.960 --> 17:01.960 which is something you may not want to do. 17:03.960 --> 17:08.960 I hope you're amazed by my roadmap compatible with the EU. 17:09.960 --> 17:12.960 So that's a roadmap for 2025. 17:12.960 --> 17:13.960 Q1. 17:13.960 --> 17:15.960 It's a recruitment generation prototype. 17:15.960 --> 17:18.960 As I said, the last PR was done yesterday. 17:18.960 --> 17:21.960 So normally everything is almost ready on data very side. 17:21.960 --> 17:24.960 So it should be complete in time. 17:24.960 --> 17:26.960 Second quarter. 17:26.960 --> 17:29.960 We'll finalize the transformation connectors. 17:29.960 --> 17:33.960 Okay, the transformation connectors are for the many full CF pipelines. 17:33.960 --> 17:35.960 So it will allow you to easily say, 17:35.960 --> 17:37.960 I want to have classification. 17:37.960 --> 17:39.960 I want to have entity extraction. 17:39.960 --> 17:40.960 I want to have translation. 17:40.960 --> 17:44.960 This type of actions when you do Vintexing phase. 17:44.960 --> 17:48.960 Second quarter as well, we provide Victor Search instead of BM25. 17:48.960 --> 17:49.960 So for the results list. 17:49.960 --> 17:54.960 So you will be able to choose either the good old BM25 or Victor Search. 17:54.960 --> 17:57.960 And then in Q3, you will have the hybrid search. 17:57.960 --> 17:59.960 Are you familiar with hybrid search? 17:59.960 --> 18:00.960 Yeah. 18:00.960 --> 18:02.960 Well, it's good. 18:02.960 --> 18:03.960 Okay, good. 18:03.960 --> 18:05.960 But the hybrid search is you do it with two queries. 18:05.960 --> 18:06.960 You do a BM25. 18:06.960 --> 18:07.960 So the base query. 18:07.960 --> 18:10.960 At the same time, you do a Victor Search query. 18:10.960 --> 18:15.960 And you merge the results into just one results set. 18:15.960 --> 18:17.960 And we will be using RRF. 18:17.960 --> 18:18.960 Reciprocare rank. 18:18.960 --> 18:19.960 Fusion. 18:19.960 --> 18:20.960 So it's always hard to say. 18:20.960 --> 18:22.960 Recipro. 18:22.960 --> 18:23.960 Okay. 18:23.960 --> 18:24.960 RRF. 18:24.960 --> 18:25.960 As an algorithm. 18:25.960 --> 18:27.960 And. 18:27.960 --> 18:29.960 Six minutes for questions. 18:29.960 --> 18:31.960 Because you're all eager to ask questions. 18:31.960 --> 18:34.960 You can go on our website for the downloads. 18:34.960 --> 18:35.960 I have already. 18:35.960 --> 18:36.960 So it's Apache license. 18:36.960 --> 18:37.960 Okay. 18:37.960 --> 18:38.960 In Java. 18:38.960 --> 18:40.960 We are not allergic to Java. 18:40.960 --> 18:42.960 If you are a doctor. 18:42.960 --> 18:43.960 Lovers. 18:43.960 --> 18:44.960 You can use the container on the curb. 18:44.960 --> 18:47.960 That's probably the quickest way for you to play with it. 18:47.960 --> 18:49.960 Lots of documentation. 18:49.960 --> 18:50.960 Okay. 18:50.960 --> 18:51.960 We are now. 18:51.960 --> 18:53.960 I think we have 300 pages of the group. 18:53.960 --> 18:55.960 Some of them are maybe. 18:55.960 --> 18:56.960 Deplicated. 18:56.960 --> 18:57.960 But we are not aware of. 18:57.960 --> 18:59.960 So if you end up with issues. 18:59.960 --> 19:01.960 But best is probably to ask us. 19:01.960 --> 19:04.960 You know, you do a GitHub discussion or GitHub issue. 19:04.960 --> 19:05.960 And you say. 19:05.960 --> 19:06.960 Is it? 19:06.960 --> 19:07.960 It's a documentation. 19:07.960 --> 19:08.960 Really. 19:08.960 --> 19:09.960 Okay. 19:09.960 --> 19:10.960 We don't know best. 19:10.960 --> 19:11.960 But it's not easy. 19:11.960 --> 19:13.960 So the code is normally on our GitHub. 19:13.960 --> 19:15.960 But we have a mirror on GitHub. 19:15.960 --> 19:18.960 But all the projects are on GitHub. 19:18.960 --> 19:19.960 And yeah. 19:19.960 --> 19:20.960 The main one is on GitHub. 19:20.960 --> 19:22.960 And the LinkedIn account. 19:22.960 --> 19:24.960 You want to follow our amazing activities. 19:24.960 --> 19:27.960 Because we just recruited the marketing and. 19:28.960 --> 19:31.960 So this day we have lots of LinkedIn contributions. 19:31.960 --> 19:32.960 So if you want to follow. 19:32.960 --> 19:33.960 That's what it was. 19:33.960 --> 19:34.960 Six months. 19:34.960 --> 19:35.960 It should follow. 19:35.960 --> 19:36.960 The. 19:36.960 --> 19:37.960 And I'm done. 19:37.960 --> 19:39.960 Thank you very much. 19:45.960 --> 19:46.960 Thank you very much. 19:46.960 --> 19:47.960 The trick. 19:47.960 --> 19:48.960 Do we have any question? 19:48.960 --> 19:49.960 Five minutes. 19:49.960 --> 19:50.960 Three. 19:50.960 --> 19:51.960 Two. 19:51.960 --> 19:52.960 Two. 19:52.960 --> 19:54.960 Two. 19:55.960 --> 19:56.960 Do you have a connector for next. 19:56.960 --> 19:58.960 Or import the documents? 19:58.960 --> 19:59.960 Uh, no. 19:59.960 --> 20:01.960 In a sense that it's it's easy. 20:01.960 --> 20:03.960 I mean, to have one without permission. 20:03.960 --> 20:04.960 It's easy because you just need to. 20:04.960 --> 20:05.960 You create an account. 20:05.960 --> 20:06.960 Next. 20:06.960 --> 20:07.960 I can't with full access to everything. 20:07.960 --> 20:08.960 And you. 20:08.960 --> 20:09.960 You index with mirror. 20:09.960 --> 20:10.960 Or which is local. 20:10.960 --> 20:14.960 Uh, but for doing it with the access control. 20:14.960 --> 20:16.960 So we asked to. 20:16.960 --> 20:17.960 A company specialized. 20:17.960 --> 20:19.960 It would not be expensive to do. 20:19.960 --> 20:20.960 But we don't have the money. 20:20.960 --> 20:21.960 For that. 20:21.960 --> 20:22.960 So. 20:22.960 --> 20:23.960 Yeah. 20:23.960 --> 20:24.960 It's doable. 20:24.960 --> 20:25.960 For release. 20:25.960 --> 20:26.960 10,000 euros. 20:26.960 --> 20:27.960 But not enough. 20:27.960 --> 20:28.960 But yeah. 20:28.960 --> 20:29.960 What would be. 20:29.960 --> 20:30.960 Thank you. 20:30.960 --> 20:32.960 Other questions. 20:32.960 --> 20:33.960 Um. 20:33.960 --> 20:34.960 Um. 20:34.960 --> 20:35.960 Is the. 20:35.960 --> 20:37.960 The index only updated. 20:37.960 --> 20:39.960 Uh, through crawling or is there. 20:39.960 --> 20:40.960 Some. 20:40.960 --> 20:41.960 So. 20:41.960 --> 20:42.960 And the passive way or is there some. 20:42.960 --> 20:45.960 Um, active way like if a document changes in one source. 20:45.960 --> 20:50.960 The source can notify the search engine to re-intex it. 20:50.960 --> 20:54.960 I mean, if it's here, uh, initially, it was thought to be both push and pull. 20:54.960 --> 20:57.960 Uh, they have implemented just the interface for the push. 20:57.960 --> 20:58.960 Okay. 20:58.960 --> 20:59.960 So for now, it's put. 20:59.960 --> 21:01.960 So we fetch the data regularly. 21:01.960 --> 21:02.960 We've crawl. 21:02.960 --> 21:03.960 Prior to crawl. 21:03.960 --> 21:04.960 So you can configure. 21:04.960 --> 21:05.960 Uh, but there is no. 21:05.960 --> 21:06.960 It. 21:06.960 --> 21:07.960 And you always crawl. 21:07.960 --> 21:08.960 Everything. 21:08.960 --> 21:09.960 I don't know. 21:09.960 --> 21:10.960 This depends on the API. 21:10.960 --> 21:11.960 So. 21:11.960 --> 21:12.960 Uh, five shares. 21:12.960 --> 21:13.960 For instance, yes. 21:13.960 --> 21:14.960 It grows everything. 21:14.960 --> 21:16.960 Although we have an agent. 21:16.960 --> 21:17.960 Uh, for. 21:17.960 --> 21:20.960 And, uh, as some be shares on windows, but monitors at the low level. 21:20.960 --> 21:22.960 Uh, the OS level, modifications. 21:22.960 --> 21:25.960 We saw those to do real delta in five shares. 21:25.960 --> 21:27.960 But for instance, for excuse. 21:27.960 --> 21:29.960 Uh, which is an amazing witty. 21:29.960 --> 21:31.960 Uh, the API gives us the delta. 21:31.960 --> 21:32.960 So it's really fast. 21:32.960 --> 21:35.960 We don't need to record everything. 21:35.960 --> 21:37.960 Um, do you have a strategy? 21:37.960 --> 21:39.960 How to deal with images and. 21:39.960 --> 21:40.960 And, uh, a lot of them. 21:40.960 --> 21:41.960 Why? 21:41.960 --> 21:44.960 I mean, let me refresh that. 21:44.960 --> 21:48.960 I mean, first, there's seem to be several strategies. 21:48.960 --> 21:49.960 How to index images. 21:49.960 --> 21:51.960 So do we have any plans for that? 21:51.960 --> 21:55.960 And the second thing is you could also consider images in the content. 21:55.960 --> 21:57.960 Uh, for the chat. 21:57.960 --> 21:59.960 So something you plan or you do. 21:59.960 --> 22:00.960 Uh, yeah. 22:00.960 --> 22:02.960 You mean as an input like we use a put to an image in the chat? 22:02.960 --> 22:06.960 No. I mean, if the content for example references an image or if the documentary index 22:06.960 --> 22:10.960 contents in image, you could provide the image to the LM. 22:10.960 --> 22:11.960 I mean, we have. 22:11.960 --> 22:12.960 Yeah. 22:12.960 --> 22:16.960 So that's that's way further than the 2025 or the map. 22:16.960 --> 22:18.960 It's how to handle the multi-modality. 22:18.960 --> 22:22.960 Uh, so we we have looked very briefly at Colpalee, for instance, 22:22.960 --> 22:25.960 for the followers of you for the multi-modality models. 22:25.960 --> 22:30.960 Um, but yeah, the graph, the holy grail of indexing this day is how to be able to really 22:30.960 --> 22:33.960 properly manage multi-modal documents. 22:33.960 --> 22:38.960 Like a document which has a graph, an image, text, formula, etc. 22:38.960 --> 22:40.960 There are ways for now. 22:41.960 --> 22:44.960 But considering where we are, it will take some time. 22:44.960 --> 22:48.960 But yeah, if you have a first pass when you do indexing, where you take everything as an image, 22:48.960 --> 22:51.960 basically, when it's looking at a text, everything is an image. 22:51.960 --> 22:55.960 And you have one LLM dedicated to detecting the types of modalities. 22:55.960 --> 22:58.960 Okay. And he says, okay, here very similar. 22:58.960 --> 23:01.960 And he calls an LLM dedicated to analyzing formulas. 23:01.960 --> 23:06.960 It stores as a metadata, a summary of what it has analyzed. 23:06.960 --> 23:08.960 So like this formula talks about this and that. 23:08.960 --> 23:10.960 This image talks about the cat, etc. 23:10.960 --> 23:12.960 And we store all that. 23:12.960 --> 23:16.960 And then when the user asks a query, the text query. 23:16.960 --> 23:18.960 So it's in the future. 23:18.960 --> 23:20.960 What the user will ask a query. 23:20.960 --> 23:23.960 We will look in this text description of the different modalities. 23:23.960 --> 23:26.960 Which ones would satisfy what the user asks? 23:26.960 --> 23:27.960 Okay. 23:27.960 --> 23:30.960 And then it will call for proper LLM specialized. 23:30.960 --> 23:35.960 So let's say I have a document where I have my financial statements as a table for the last two 23:35.960 --> 23:40.960 years. And the question is, what is the financial revenue expected for the next two years? 23:40.960 --> 23:41.960 Obviously. 23:41.960 --> 23:42.960 Obviously. 23:42.960 --> 23:48.960 The dream is that the LLM, the system will choose the LLM specialized in the graph analysis. 23:48.960 --> 23:51.960 And then do the computation for the forecast because obviously this is where it stands. 23:51.960 --> 23:53.960 So what's the goal? 23:53.960 --> 23:58.960 The pieces are more or less ver, but I heard that Colpally generates so many vectors for one single page. 23:58.960 --> 24:01.960 But it would compute in this row, it was calamity for now. 24:01.960 --> 24:03.960 So it's ongoing work. 24:03.960 --> 24:05.960 But not there yet. 24:05.960 --> 24:06.960 Thank you. 24:06.960 --> 24:08.960 Any other question? 24:08.960 --> 24:11.960 We have time for one more? 24:11.960 --> 24:12.960 No. 24:12.960 --> 24:13.960 Okay. 24:13.960 --> 24:14.960 And thank you very much. 24:14.960 --> 24:15.960 Thank you very much.