WEBVTT captioned by sachac NOTE Search in daily workflows 00:00:00.000 --> 00:00:03.399 Hello, my name is Zachary Romero, and today I'll be going 00:00:03.400 --> 00:00:08.115 over p-search, a local search engine in Emacs. 00:00:08.116 --> 00:00:12.398 Search these days is everywhere in software, from text editors, 00:00:12.399 --> 00:00:18.359 to IDEs, to most online websites. These tools tend to fall 00:00:18.360 --> 00:00:25.839 into one of two categories. One are tools that run locally, 00:00:25.840 --> 00:00:31.279 and work by matching string to text. The most common 00:00:31.280 --> 00:00:35.639 example of this is grep. In Emacs, there are a lot of 00:00:35.640 --> 00:00:38.959 extensions which provide functionality on top of these 00:00:38.960 --> 00:00:42.388 tools, such as projectile-grep, deadgrep, 00:00:42.389 --> 00:00:46.849 consult-ripgrep. Most editors have some sort of 00:00:46.850 --> 00:00:52.691 search current project feature. Most of the time, 00:00:52.692 --> 00:00:56.393 some of these tools have features like regular expressions, 00:00:56.394 --> 00:00:59.215 or you can specify file extension, 00:00:59.216 --> 00:01:01.636 or a directory you want to search in, 00:01:01.637 --> 00:01:03.957 but features are pretty limited. 00:01:03.958 --> 00:01:07.919 The other kind of search we use are usually hosted online, 00:01:07.920 --> 00:01:12.302 and they usually search a vast corpus of data. 00:01:12.303 --> 00:01:15.639 These are usually proprietary 00:01:15.640 --> 00:01:18.765 online services such as Google, GitHub, 00:01:18.766 --> 00:01:24.199 SourceGraph for code. NOTE Problems with editor search tools 00:01:24.200 --> 00:01:28.839 The kind of search feature that editors 00:01:28.840 --> 00:01:36.719 usually have have a lot of downsides to them. For one, a lot 00:01:36.720 --> 00:01:38.839 of times you don't know the exact search string you're 00:01:38.840 --> 00:01:42.783 searching for. Some complicated term like this 00:01:42.784 --> 00:01:46.860 high volume demand partner, you know, do you know if... 00:01:46.861 --> 00:01:49.708 Are some words abbreviated, is it capitalized, 00:01:49.709 --> 00:01:53.089 is it in kebab case, camel case, snake case? 00:01:53.090 --> 00:01:57.571 You often have to search all these variations. 00:01:57.572 --> 00:02:05.434 Another downside is that the search results returned 00:02:05.435 --> 00:02:07.769 contain a lot of noise. For example, 00:02:07.770 --> 00:02:10.816 you may get a lot of test files. 00:02:10.817 --> 00:02:13.537 If the tool hits your vendor directory, 00:02:13.538 --> 00:02:17.199 it may get a bunch of results from libraries 00:02:17.200 --> 00:02:22.879 you're using, which most are not helpful. Another downside 00:02:22.880 --> 00:02:26.679 is that the order given is, well, there's no meaning to the 00:02:26.680 --> 00:02:30.319 order. It's usually just the search order that the tool 00:02:30.320 --> 00:02:34.639 happens to look in first. 00:02:34.640 --> 00:02:38.639 Another thing is, so when you're searching, you oftentimes 00:02:38.640 --> 00:02:41.639 have to keep the state of the searches in your head. For 00:02:41.640 --> 00:02:46.639 example, you try one search, you see the results, find the 00:02:46.640 --> 00:02:49.639 results you think are relevant, keep them in your head, run 00:02:49.640 --> 00:02:52.519 search number two, look through the results, kind of 00:02:52.520 --> 00:02:56.119 combine these different search results in your head until 00:02:56.120 --> 00:02:59.970 you get an idea of which ones might be relevant. 00:02:59.971 --> 00:03:04.515 Another thing is that the search primitives are fairly limited. 00:03:04.516 --> 00:03:10.599 So yeah, you can search regular expressions, but you can't 00:03:10.600 --> 00:03:14.719 really define complex things like, I want to search files in 00:03:14.720 --> 00:03:18.439 this directory, and this directory, and this directory, 00:03:18.440 --> 00:03:22.319 except these subdirectories, and accept test files, and I 00:03:22.320 --> 00:03:25.559 only want files with this file extension. Criteria like 00:03:25.560 --> 00:03:28.919 that are really hard to... I'm sure they're possible in tools 00:03:28.920 --> 00:03:34.479 like grep, but they're pretty hard to construct. 00:03:34.480 --> 00:03:38.199 And lastly, there's no notion of any relevance. All the 00:03:38.200 --> 00:03:42.039 results you get back, I mean, you don't know, is the search 00:03:42.040 --> 00:03:43.095 more relevant? Is it twice as relevant? Is it 00:03:43.096 --> 00:03:52.279 100 times more relevant? These tools usually don't provide 00:03:52.280 --> 00:03:58.232 such information. NOTE Information retrieval 00:03:58.233 --> 00:04:00.394 There's a field called information retrieval, 00:04:00.395 --> 00:04:02.616 and this deals with this exact problem. 00:04:02.617 --> 00:04:04.718 You have lots of data you're searching for. 00:04:04.719 --> 00:04:09.261 How do you construct a search query? 00:04:09.262 --> 00:04:09.839 How do you get results back fast? How do you 00:04:09.840 --> 00:04:14.519 rank which ones are most relevant? How do you evaluate 00:04:14.520 --> 00:04:20.079 your search system to see if it's getting better or worse? 00:04:20.080 --> 00:04:23.119 There's a lot of work, a lot of books written on the topic of 00:04:23.120 --> 00:04:28.159 information retrieval. If one wants to improve 00:04:28.160 --> 00:04:31.879 searching in Emacs, then drawing inspiration from this 00:04:31.880 --> 00:04:34.295 field is necessary. NOTE Search engine in Emacs: the index 00:04:34.296 --> 00:04:41.383 The first aspect of information retrieval is the index. 00:04:41.384 --> 00:04:46.608 The reverse index is what search engines use to find results really fast. 00:04:46.609 --> 00:04:51.454 Essentially, it's a map of search term 00:04:51.455 --> 00:04:54.738 to locations where that term is located. 00:04:54.739 --> 00:04:57.079 You'll have all the terms or maybe even parts of 00:04:57.080 --> 00:04:59.159 the terms, and then you'll have all the locations where 00:04:59.160 --> 00:05:02.119 they're located. Any query could easily look up 00:05:02.120 --> 00:05:05.919 where things are located, join results together, and 00:05:05.920 --> 00:05:12.879 that's how they get the results to be really fast. For this 00:05:12.880 --> 00:05:19.159 project, I decided to forgo creating an index altogether. 00:05:19.160 --> 00:05:23.759 An index is pretty complicated to maintain because 00:05:23.760 --> 00:05:27.319 it always has to be in sync. Any time you open a file and save 00:05:27.320 --> 00:05:29.959 it, you would have to re-index, you would have to make sure 00:05:29.960 --> 00:05:32.559 that file is re-indexed properly. Then you have the 00:05:32.560 --> 00:05:36.119 whole issue of, well, if you're searching in Emacs, 00:05:36.120 --> 00:05:38.799 you have all these projects, this directory, 00:05:38.800 --> 00:05:42.479 that directory, how do you know which? Do you always have to 00:05:42.480 --> 00:05:47.399 keep them in sync? It's quite a hard task to handle 00:05:47.400 --> 00:05:53.079 that. Then on the other end, tools like ripgrep can 00:05:53.080 --> 00:05:59.119 search very fast. Even though they can't search maybe on the 00:05:59.120 --> 00:06:03.919 order of tens of thousands of repositories, for a local 00:06:03.920 --> 00:06:06.039 setting, they should be plenty fast enough. 00:06:06.040 --> 00:06:12.239 I benchmarked. Ripgrep, for example, is 00:06:12.240 --> 00:06:15.959 on the order of gigabytes per second. 00:06:15.960 --> 00:06:19.239 Definitely, it can search a few pretty big size 00:06:19.240 --> 00:06:21.756 repositories. NOTE Search engine in Emacs: Ranking 00:06:21.757 --> 00:06:24.799 Next main task. We decided not to use an 00:06:24.800 --> 00:06:29.959 index. Next task is how do we rank search results? So there's 00:06:29.960 --> 00:06:33.439 two main algorithms that are used these days. The first 00:06:33.440 --> 00:06:36.519 one is tf-idf, which stands for term frequency, inverse 00:06:36.520 --> 00:06:43.039 target frequency. Then there's BM25, which is sort of a 00:06:43.040 --> 00:06:43.552 modified tf-idf algorithm. NOTE tf-idf: term-frequency x inverse-document-frequency 00:06:43.553 --> 00:06:45.679 tf-idf, without going into 00:06:45.680 --> 00:06:49.159 too much detail, essentially multiplies two terms. One 00:06:49.160 --> 00:06:51.879 is the term frequency, and then you multiply it by the 00:06:51.880 --> 00:06:54.559 inverse document frequency. The term frequency is a 00:06:54.560 --> 00:06:58.519 measure of how often that search term occurs. The 00:06:58.520 --> 00:07:00.799 inverse document frequency is a measure of how much 00:07:00.800 --> 00:07:06.199 information that term provides. If the term occurs a lot, 00:07:06.200 --> 00:07:08.719 then it gets a higher score in the term frequency section. 00:07:08.720 --> 00:07:12.399 But if it's a common word that exists in a lot of documents, 00:07:12.400 --> 00:07:13.900 then its inverse document frequency goes down. 00:07:13.901 --> 00:07:20.879 It kind of scores it less. You'll find that words like the, 00:07:20.880 --> 00:07:25.959 in, is, these really common words, since they occur 00:07:25.960 --> 00:07:29.199 everywhere, their inverse document frequency is 00:07:29.200 --> 00:07:32.479 essentially zero. They don't really count towards a 00:07:32.480 --> 00:07:35.679 score. But when you have rare words that only occur in a 00:07:35.680 --> 00:07:37.679 few documents, they're weighted a lot more. So the more 00:07:37.680 --> 00:07:41.159 those rare words occur, they boost the score higher. NOTE BM25 00:07:41.160 --> 00:07:48.839 BM25 is a modification of this. It's essentially TF, it's 00:07:48.840 --> 00:07:53.119 essentially the previous one, except it dampens out terms 00:07:53.120 --> 00:07:55.439 that occur more often. Imagine you have a bunch of 00:07:55.440 --> 00:07:59.359 documents. One has a term 10 times, one has a term, that same 00:07:59.360 --> 00:08:02.439 term a hundred times, another has a thousand times. 00:08:02.440 --> 00:08:06.799 You'll see the score dampens off as the number of 00:08:06.800 --> 00:08:10.639 occurrences increases. That prevents any one term from 00:08:10.640 --> 00:08:16.559 overpowering the score. This is the algorithm I ended up 00:08:16.560 --> 00:08:21.039 choosing for my implementation. So with a plan of using a 00:08:21.040 --> 00:08:29.559 command line tool like ripgrep to get term occurrences, and 00:08:29.560 --> 00:08:36.799 then using a scoring algorithm like BM25 to rank the terms, 00:08:36.800 --> 00:08:40.079 we can combine this together and create a simple search 00:08:40.080 --> 00:08:41.199 mechanism. NOTE Searching with p-search 00:08:41.200 --> 00:08:47.439 Here we're in the directory for the Emacs source code. 00:08:47.440 --> 00:08:53.479 Let's say we want to search for the display code. We 00:08:53.480 --> 00:08:58.679 run the p-search command, starting the search engine. It 00:08:58.680 --> 00:09:01.159 opens up. We notice it has three sections, the candidate 00:09:01.160 --> 00:09:05.199 generators, the priors, and the search results. The 00:09:05.200 --> 00:09:09.999 candidate generators generates the search space we're 00:09:10.000 --> 00:09:14.719 looking on. These are all composable and you can add as 00:09:14.720 --> 00:09:19.719 many as you want. So with this, it specifies that here 00:09:19.720 --> 00:09:25.239 we're searching on the file system and we're searching in 00:09:25.240 --> 00:09:30.799 this directory. We're using the ripgrep tool to search 00:09:30.800 --> 00:09:33.359 with, and we want to make sure that we're searching only on 00:09:33.360 --> 00:09:40.479 files committed to Git. Here we see the search results. 00:09:40.480 --> 00:09:45.159 Notice here is their final probability. Here, notice 00:09:45.160 --> 00:09:47.079 that they're all the same, and they're the same because we 00:09:47.080 --> 00:09:50.719 don't have any search criteria specified here. Suppose 00:09:50.720 --> 00:09:55.679 we want to search for display-related code. We add a 00:09:55.680 --> 00:09:57.359 query: display. 00:09:57.360 --> 00:10:06.559 So then it spins off the processes, gets the search term 00:10:06.560 --> 00:10:10.879 counts and calculates the new scores. Notice here that 00:10:10.880 --> 00:10:15.759 the results that come on top are just at first glance appear 00:10:15.760 --> 00:10:19.919 to be relevant to display. Remember, if we compare 00:10:19.920 --> 00:10:25.079 that to just running a ripgrep raw, notice here we're 00:10:25.080 --> 00:10:31.279 getting 53,000 results and it's pretty hard to go through 00:10:31.280 --> 00:10:34.319 these results and make sense of it. 00:10:34.320 --> 00:10:41.456 So that's p-search in a nutshell. NOTE Flight AF 447 00:10:41.457 --> 00:10:45.982 Next, I wanted to talk about the story of Flight 447. 00:10:45.983 --> 00:10:49.326 Flight 447 going from Rio de Janeiro to Paris 00:10:49.327 --> 00:10:51.509 crashed somewhere in the Atlantic Ocean 00:10:51.510 --> 00:10:54.713 on June 1st, 2009, killing everyone on board. 00:10:54.714 --> 00:10:56.894 Four search attempts were made to find the wreckage. 00:10:56.895 --> 00:11:01.075 None of them were successful, except the finding of some debris 00:11:01.076 --> 00:11:05.479 and a dead body. It was decided that they really wanted 00:11:05.480 --> 00:11:09.519 to find the wreckage to retrieve data as to why the search 00:11:09.520 --> 00:11:14.639 occurred. This occurred two years after the 00:11:14.640 --> 00:11:19.959 initial crash. With this next search attempt, they 00:11:19.960 --> 00:11:23.199 wanted to create a probability distribution of where the 00:11:23.200 --> 00:11:26.759 crash could be. The only piece of concrete data they had 00:11:26.760 --> 00:11:35.079 was a GPS signal from the ship at 210 containing the GPS 00:11:35.080 --> 00:11:40.239 location of the plane was at 2.98 degrees north, 30.59 00:11:40.240 --> 00:11:44.719 degrees west. That was the only data they had to go off of. 00:11:44.720 --> 00:11:50.079 So they drew a circle around that point 00:11:50.080 --> 00:11:54.679 with a radius of 40 nautical miles. They assumed that 00:11:54.680 --> 00:11:57.479 anything outside the circle would have been impossible for 00:11:57.480 --> 00:12:01.239 the ship to reach. This was the starting point for 00:12:01.240 --> 00:12:04.799 creating the probability distribution of where the 00:12:04.800 --> 00:12:08.119 wreckage occurred. Anything outside the circle, they 00:12:08.120 --> 00:12:09.639 assumed it was impossible to reach. 00:12:09.640 --> 00:12:16.479 The only other pieces of data were the four failed search 00:12:16.480 --> 00:12:21.719 attempts and then some of the debris found. One thing they 00:12:21.720 --> 00:12:26.159 did decide was to look at similar crashes where control was 00:12:26.160 --> 00:12:30.319 lost to analyze where the crashes landed, compared to where 00:12:30.320 --> 00:12:37.399 the loss of control started. This probability 00:12:37.400 --> 00:12:43.479 distribution, the circular normal distribution was 00:12:43.480 --> 00:12:47.919 decided upon. Here you can see that the center has a lot 00:12:47.920 --> 00:12:51.879 higher chance of finding the wreckage. As you go away 00:12:51.880 --> 00:12:55.399 from the center, the probability of finding the wreckage 00:12:55.400 --> 00:13:02.319 decreases a lot. The next thing they looked at was, well, 00:13:02.320 --> 00:13:05.959 they noticed they had retrieved some dead bodies from the 00:13:05.960 --> 00:13:12.959 wreckage. So they thought that they could calculate the 00:13:12.960 --> 00:13:18.439 backward drift on that particular day to find where the 00:13:18.440 --> 00:13:21.479 crash might've occurred. If they found bodies at a 00:13:21.480 --> 00:13:25.119 particular location, they can kind of work backwards from 00:13:25.120 --> 00:13:30.665 that in order to find where the initial crash occurred. 00:13:30.666 --> 00:13:34.719 So here you can see the probability distribution based off of 00:13:34.720 --> 00:13:40.279 the backward drift model. Here you see the darker colors 00:13:40.280 --> 00:13:46.159 have a higher probability of finding the location. So 00:13:46.160 --> 00:13:50.679 with all these pieces of data, so with that circular 40 00:13:50.680 --> 00:13:54.959 nautical mile uniform distribution, with that circular 00:13:54.960 --> 00:14:02.199 normal distribution of comparing similar crashes, as well 00:14:02.200 --> 00:14:07.439 as with the backward drift, they were able to combine all 00:14:07.440 --> 00:14:08.559 three of these pieces 00:14:08.560 --> 00:14:14.599 in order to come up with a final prior distribution of where 00:14:14.600 --> 00:14:19.519 the wreckage occurred. So this is what the final model 00:14:19.520 --> 00:14:24.719 they came upon. Here you can see it has that 40 nautical 00:14:24.720 --> 00:14:29.679 mile radius circle. It has that darker center, which 00:14:29.680 --> 00:14:32.039 indicates a higher probability because of the 00:14:32.040 --> 00:14:38.959 crash similarity. Then here you also see along this line 00:14:38.960 --> 00:14:50.799 has a slightly higher probability due to the backward drift 00:14:50.800 --> 00:14:52.119 distribution. 00:14:52.120 --> 00:14:56.559 So the next thing is, since they had performed searches, 00:14:56.560 --> 00:15:00.559 they decided to incorporate the data from those searches 00:15:00.560 --> 00:15:04.759 into their new distribution. Here you can see places 00:15:04.760 --> 00:15:08.879 where they searched initially. If you think about it, 00:15:08.880 --> 00:15:11.399 you can assume that, well, if you search for something, 00:15:11.400 --> 00:15:14.199 there's a good chance you'll find it, but not necessarily. 00:15:14.200 --> 00:15:18.439 Anywhere where they searched, the probability of it 00:15:18.440 --> 00:15:22.839 finding it there is greatly reduced. It's not zero because 00:15:22.840 --> 00:15:26.879 obviously you can look for something and miss it, but it kind 00:15:26.880 --> 00:15:31.119 of reduces the probability that we would expect to find it in 00:15:31.120 --> 00:15:36.679 those already searched locations. This is the 00:15:36.680 --> 00:15:41.919 posterior distribution or distribution after counting 00:15:41.920 --> 00:15:44.559 observations made. 00:15:44.560 --> 00:15:48.759 Here we can see kind of these cutouts of where the 00:15:48.760 --> 00:15:53.959 previous searches occurred. This is the final 00:15:53.960 --> 00:15:56.999 distribution they went off of to perform the subsequent 00:15:57.000 --> 00:16:01.999 search. In the end, the wreckage was found at a point close to 00:16:02.000 --> 00:16:06.770 the center here, thus validating this methodology. NOTE Modifying priors 00:16:06.771 --> 00:16:10.332 We can see the power of this Bayesian search methodology 00:16:10.333 --> 00:16:13.999 in the way that we could take information from all the sources we had. 00:16:14.000 --> 00:16:19.237 We could draw analogies to similar situations. 00:16:19.238 --> 00:16:22.479 We can quantify these, combine them into a model, 00:16:22.480 --> 00:16:27.893 and then also update our model according to each observation we make. 00:16:27.894 --> 00:16:30.359 I think there's a lot of similarities to be drawn with 00:16:30.360 --> 00:16:35.159 searching on a computer in the sense that when we search for 00:16:35.160 --> 00:16:39.399 something, there's oftentimes a story we kind of have as to 00:16:39.400 --> 00:16:43.959 what search terms exist, where we expect to find the file. 00:16:43.960 --> 00:16:46.719 For example, if you're implementing a new feature, you'll 00:16:46.720 --> 00:16:49.919 often have some search terms in mind that you think will be 00:16:49.920 --> 00:16:54.719 relevant. Some search terms, you might think they have a 00:16:54.720 --> 00:16:57.599 possibility of being relevant, but maybe you're not sure. 00:16:57.600 --> 00:17:02.879 There's some directories where you know that they're not 00:17:02.880 --> 00:17:07.759 relevant. There's other criteria like, well, you know that 00:17:07.760 --> 00:17:11.399 maybe somebody in particular worked on this code. 00:17:11.400 --> 00:17:16.319 What if you could incorporate that information? Like, I know 00:17:16.320 --> 00:17:21.399 this author, he's always working on this feature. What if 00:17:21.400 --> 00:17:25.519 I just give the files that this person works on a higher 00:17:25.520 --> 00:17:32.599 probability than ones he doesn't work on? Or maybe you think 00:17:32.600 --> 00:17:38.599 that this is a file that's committed too often. You think 00:17:38.600 --> 00:17:43.439 that maybe the amount of times of commits it receives 00:17:43.440 --> 00:17:47.719 should change your probability of this file being 00:17:47.720 --> 00:17:52.839 relevant. That's where p-search comes in. 00:17:52.840 --> 00:17:57.679 Its aim is to be a framework in order to incorporate all these 00:17:57.680 --> 00:18:01.359 sorts of different prior information into your searching 00:18:01.360 --> 00:18:05.999 process. You're able to say things like, I want files 00:18:06.000 --> 00:18:11.119 authored by this user to be given higher probability. I want 00:18:11.120 --> 00:18:13.919 this author to be given a lower priority. I know this author 00:18:13.920 --> 00:18:18.759 never works on this code. If he has a commit, then lower its 00:18:18.760 --> 00:18:24.679 probability, or you can specify specific paths, or you can 00:18:24.680 --> 00:18:30.199 specify multiple search terms, weighing different ones 00:18:30.200 --> 00:18:38.919 according to how you think those terms should be relevant. 00:18:38.920 --> 00:18:42.079 So with p-search, we're able to incorporate information 00:18:42.080 --> 00:18:46.279 from multiple sources. Here, for example, we have a prior 00:18:46.280 --> 00:18:52.079 of type git author, and we're looking for all of the files 00:18:52.080 --> 00:18:56.719 that are committed to by Lars. So the more commits he has, 00:18:56.720 --> 00:19:01.399 the higher probability is given to that file. Suppose 00:19:01.400 --> 00:19:04.559 there's a feature I know he worked on, but I don't know the 00:19:04.560 --> 00:19:09.159 file or necessarily even key terms of it. Well, with this, I 00:19:09.160 --> 00:19:12.140 can incorporate that information. 00:19:12.141 --> 00:19:15.999 So let's search again. Let's add display. 00:19:16.000 --> 00:19:22.959 Let's see what responses we get back here. We can add 00:19:22.960 --> 00:19:27.199 as many of these criteria as we want. We can even specify that 00:19:27.200 --> 00:19:31.519 the title of the file name should be a certain type. Let's 00:19:31.520 --> 00:19:36.599 say we're only concerned about C files. We add the file 00:19:36.600 --> 00:19:45.399 name should contain .c in it. With this, now we 00:19:45.400 --> 00:19:51.319 notice that all of the C files containing display authored 00:19:51.320 --> 00:19:56.279 by Lars should be given higher probability. We can 00:19:56.280 --> 00:20:02.719 continue to add these priors as we feel fit. The workflow 00:20:02.720 --> 00:20:07.519 that I found helps when searching is that you'll add 00:20:07.520 --> 00:20:11.359 criteria, you'll see some good results come up and some bad 00:20:11.360 --> 00:20:15.319 results come up. So you'll often find a pattern in those 00:20:15.320 --> 00:20:18.839 bad results, like, oh, I don't want test files, or this 00:20:18.840 --> 00:20:22.679 directory isn't relevant, or something like that. Then 00:20:22.680 --> 00:20:27.199 you can update your prior distribution, adding its 00:20:27.200 --> 00:20:31.119 criteria, and then rerun it, and then it will get different 00:20:31.120 --> 00:20:35.159 probabilities for the files. So in the end, you'll have a 00:20:35.160 --> 00:20:37.639 list of results that's tailor-made to the thing you're 00:20:37.640 --> 00:20:40.404 searching for. NOTE Importance 00:20:40.405 --> 00:20:41.639 There's a couple of other features I 00:20:41.640 --> 00:20:49.079 want to go through. One thing is that each of these priors, 00:20:49.080 --> 00:20:55.839 you can specify the importance. In other words, how 00:20:55.840 --> 00:21:01.119 important is this particular piece of information to your 00:21:01.120 --> 00:21:05.199 search? So here, everything is of importance medium. But 00:21:05.200 --> 00:21:07.879 let's say I really care about something having the word 00:21:07.880 --> 00:21:12.679 display in it. I'm going to change its importance. 00:21:12.680 --> 00:21:18.599 Instead of medium, I'll change its importance to high. 00:21:18.600 --> 00:21:23.279 What that does essentially is things that don't have 00:21:23.280 --> 00:21:28.079 display in it are given a much bigger penalty and things with 00:21:28.080 --> 00:21:28.128 the word display in it are rated much higher. 00:21:28.129 --> 00:21:38.559 With this, we're able to fine-tune the results that we get. NOTE Complement or inverse 00:21:38.560 --> 00:21:45.639 Another thing you can do is that you can add the complement or 00:21:45.640 --> 00:21:49.759 the inverse of certain queries. Let's say you want to 00:21:49.760 --> 00:21:53.239 search for display, but you don't want it to contain the word 00:21:53.240 --> 00:21:58.039 frame. With the complement option on, when we create this 00:21:58.040 --> 00:22:01.839 search prior, now it's going to be searching for frame, but 00:22:01.840 --> 00:22:04.959 instead of increasing the search score, it's going to 00:22:04.960 --> 00:22:06.999 decrease it if it contains the word frame. 00:22:07.000 --> 00:22:14.319 So here, things related to frame are kind of 00:22:14.320 --> 00:22:18.079 deprioritized. We can also say that we really don't want 00:22:18.080 --> 00:22:21.599 the search to contain the word frame by increasing its 00:22:21.600 --> 00:22:27.199 importance. So with all these composable pieces, we can 00:22:27.200 --> 00:22:33.412 create kind of a search that's tailor-made to our needs. 00:22:33.413 --> 00:22:35.759 That concludes this talk. There's a lot more I could talk 00:22:35.760 --> 00:22:37.799 about with regards to research, so definitely follow the 00:22:37.800 --> 00:22:40.639 project if you're interested. Thanks for watching, and I 00:22:40.640 --> 00:22:42.240 hope you enjoy the rest of the conference.