Transcript

[[!template text="""Hello, my name is Zachary Romero, and today I'll be going""" start="00:00:00.000" video="mainVideo-p-search" id="subtitle"]] [[!template text="""over p-search, a local search engine in Emacs.""" start="00:00:03.400" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Search these days is everywhere in software, from text editors,""" start="00:00:08.116" video="mainVideo-p-search" id="subtitle"]] [[!template text="""to IDEs, to most online websites. These tools tend to fall""" start="00:00:12.399" video="mainVideo-p-search" id="subtitle"]] [[!template text="""into one of two categories. One are tools that run locally,""" start="00:00:18.360" video="mainVideo-p-search" id="subtitle"]] [[!template text="""and work by matching string to text. The most common""" start="00:00:25.840" video="mainVideo-p-search" id="subtitle"]] [[!template text="""example of this is grep. In Emacs, there are a lot of""" start="00:00:31.280" video="mainVideo-p-search" id="subtitle"]] [[!template text="""extensions which provide functionality on top of these""" start="00:00:35.640" video="mainVideo-p-search" id="subtitle"]] [[!template text="""tools, such as projectile-grep, deadgrep,""" start="00:00:38.960" video="mainVideo-p-search" id="subtitle"]] [[!template text="""consult-ripgrep. Most editors have some sort of""" start="00:00:42.389" video="mainVideo-p-search" id="subtitle"]] [[!template text="""search current project feature. Most of the time,""" start="00:00:46.850" video="mainVideo-p-search" id="subtitle"]] [[!template text="""some of these tools have features like regular expressions,""" start="00:00:52.692" video="mainVideo-p-search" id="subtitle"]] [[!template text="""or you can specify file extension,""" start="00:00:56.394" video="mainVideo-p-search" id="subtitle"]] [[!template text="""or a directory you want to search in,""" start="00:00:59.216" video="mainVideo-p-search" id="subtitle"]] [[!template text="""but features are pretty limited.""" start="00:01:01.637" video="mainVideo-p-search" id="subtitle"]] [[!template text="""The other kind of search we use are usually hosted online,""" start="00:01:03.958" video="mainVideo-p-search" id="subtitle"]] [[!template text="""and they usually search a vast corpus of data.""" start="00:01:07.920" video="mainVideo-p-search" id="subtitle"]] [[!template text="""These are usually proprietary""" start="00:01:12.303" video="mainVideo-p-search" id="subtitle"]] [[!template text="""online services such as Google, GitHub,""" start="00:01:15.640" video="mainVideo-p-search" id="subtitle"]] [[!template text="""SourceGraph for code.""" start="00:01:18.766" video="mainVideo-p-search" id="subtitle"]] [[!template text="""The kind of search feature that editors""" start="00:01:24.200" video="mainVideo-p-search" id="subtitle"]] [[!template text="""usually have have a lot of downsides to them. For one, a lot""" start="00:01:28.840" video="mainVideo-p-search" id="subtitle"]] [[!template text="""of times you don't know the exact search string you're""" start="00:01:36.720" video="mainVideo-p-search" id="subtitle"]] [[!template text="""searching for. Some complicated term like this""" start="00:01:38.840" video="mainVideo-p-search" id="subtitle"]] [[!template text="""high volume demand partner, you know, do you know if...""" start="00:01:42.784" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Are some words abbreviated, is it capitalized,""" start="00:01:46.861" video="mainVideo-p-search" id="subtitle"]] [[!template text="""is it in kebab case, camel case, snake case?""" start="00:01:49.709" video="mainVideo-p-search" id="subtitle"]] [[!template text="""You often have to search all these variations.""" start="00:01:53.090" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Another downside is that the search results returned""" start="00:01:57.572" video="mainVideo-p-search" id="subtitle"]] [[!template text="""contain a lot of noise. For example,""" start="00:02:05.435" video="mainVideo-p-search" id="subtitle"]] [[!template text="""you may get a lot of test files.""" start="00:02:07.770" video="mainVideo-p-search" id="subtitle"]] [[!template text="""If the tool hits your vendor directory,""" start="00:02:10.817" video="mainVideo-p-search" id="subtitle"]] [[!template text="""it may get a bunch of results from libraries""" start="00:02:13.538" video="mainVideo-p-search" id="subtitle"]] [[!template text="""you're using, which most are not helpful. Another downside""" start="00:02:17.200" video="mainVideo-p-search" id="subtitle"]] [[!template text="""is that the order given is, well, there's no meaning to the""" start="00:02:22.880" video="mainVideo-p-search" id="subtitle"]] [[!template text="""order. It's usually just the search order that the tool""" start="00:02:26.680" video="mainVideo-p-search" id="subtitle"]] [[!template text="""happens to look in first.""" start="00:02:30.320" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Another thing is, so when you're searching, you oftentimes""" start="00:02:34.640" video="mainVideo-p-search" id="subtitle"]] [[!template text="""have to keep the state of the searches in your head. For""" start="00:02:38.640" video="mainVideo-p-search" id="subtitle"]] [[!template text="""example, you try one search, you see the results, find the""" start="00:02:41.640" video="mainVideo-p-search" id="subtitle"]] [[!template text="""results you think are relevant, keep them in your head, run""" start="00:02:46.640" video="mainVideo-p-search" id="subtitle"]] [[!template text="""search number two, look through the results, kind of""" start="00:02:49.640" video="mainVideo-p-search" id="subtitle"]] [[!template text="""combine these different search results in your head until""" start="00:02:52.520" video="mainVideo-p-search" id="subtitle"]] [[!template text="""you get an idea of which ones might be relevant.""" start="00:02:56.120" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Another thing is that the search primitives are fairly limited.""" start="00:02:59.971" video="mainVideo-p-search" id="subtitle"]] [[!template text="""So yeah, you can search regular expressions, but you can't""" start="00:03:04.516" video="mainVideo-p-search" id="subtitle"]] [[!template text="""really define complex things like, I want to search files in""" start="00:03:10.600" video="mainVideo-p-search" id="subtitle"]] [[!template text="""this directory, and this directory, and this directory,""" start="00:03:14.720" video="mainVideo-p-search" id="subtitle"]] [[!template text="""except these subdirectories, and accept test files, and I""" start="00:03:18.440" video="mainVideo-p-search" id="subtitle"]] [[!template text="""only want files with this file extension. Criteria like""" start="00:03:22.320" video="mainVideo-p-search" id="subtitle"]] [[!template text="""that are really hard to... I'm sure they're possible in tools""" start="00:03:25.560" video="mainVideo-p-search" id="subtitle"]] [[!template text="""like grep, but they're pretty hard to construct.""" start="00:03:28.920" video="mainVideo-p-search" id="subtitle"]] [[!template text="""And lastly, there's no notion of any relevance. All the""" start="00:03:34.480" video="mainVideo-p-search" id="subtitle"]] [[!template text="""results you get back, I mean, you don't know, is the search""" start="00:03:38.200" video="mainVideo-p-search" id="subtitle"]] [[!template text="""more relevant? Is it twice as relevant? Is it""" start="00:03:42.040" video="mainVideo-p-search" id="subtitle"]] [[!template text="""100 times more relevant? These tools usually don't provide""" start="00:03:43.096" video="mainVideo-p-search" id="subtitle"]] [[!template text="""such information.""" start="00:03:52.280" video="mainVideo-p-search" id="subtitle"]] [[!template text="""There's a field called information retrieval,""" start="00:03:58.233" video="mainVideo-p-search" id="subtitle"]] [[!template text="""and this deals with this exact problem.""" start="00:04:00.395" video="mainVideo-p-search" id="subtitle"]] [[!template text="""You have lots of data you're searching for.""" start="00:04:02.617" video="mainVideo-p-search" id="subtitle"]] [[!template text="""How do you construct a search query?""" start="00:04:04.719" video="mainVideo-p-search" id="subtitle"]] [[!template text="""How do you get results back fast? How do you""" start="00:04:09.262" video="mainVideo-p-search" id="subtitle"]] [[!template text="""rank which ones are most relevant? How do you evaluate""" start="00:04:09.840" video="mainVideo-p-search" id="subtitle"]] [[!template text="""your search system to see if it's getting better or worse?""" start="00:04:14.520" video="mainVideo-p-search" id="subtitle"]] [[!template text="""There's a lot of work, a lot of books written on the topic of""" start="00:04:20.080" video="mainVideo-p-search" id="subtitle"]] [[!template text="""information retrieval. If one wants to improve""" start="00:04:23.120" video="mainVideo-p-search" id="subtitle"]] [[!template text="""searching in Emacs, then drawing inspiration from this""" start="00:04:28.160" video="mainVideo-p-search" id="subtitle"]] [[!template text="""field is necessary.""" start="00:04:31.880" video="mainVideo-p-search" id="subtitle"]] [[!template text="""The first aspect of information retrieval is the index.""" start="00:04:34.296" video="mainVideo-p-search" id="subtitle"]] [[!template text="""The reverse index is what search engines use to find results really fast.""" start="00:04:41.384" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Essentially, it's a map of search term""" start="00:04:46.609" video="mainVideo-p-search" id="subtitle"]] [[!template text="""to locations where that term is located.""" start="00:04:51.455" video="mainVideo-p-search" id="subtitle"]] [[!template text="""You'll have all the terms or maybe even parts of""" start="00:04:54.739" video="mainVideo-p-search" id="subtitle"]] [[!template text="""the terms, and then you'll have all the locations where""" start="00:04:57.080" video="mainVideo-p-search" id="subtitle"]] [[!template text="""they're located. Any query could easily look up""" start="00:04:59.160" video="mainVideo-p-search" id="subtitle"]] [[!template text="""where things are located, join results together, and""" start="00:05:02.120" video="mainVideo-p-search" id="subtitle"]] [[!template text="""that's how they get the results to be really fast. For this""" start="00:05:05.920" video="mainVideo-p-search" id="subtitle"]] [[!template text="""project, I decided to forgo creating an index altogether.""" start="00:05:12.880" video="mainVideo-p-search" id="subtitle"]] [[!template text="""An index is pretty complicated to maintain because""" start="00:05:19.160" video="mainVideo-p-search" id="subtitle"]] [[!template text="""it always has to be in sync. Any time you open a file and save""" start="00:05:23.760" video="mainVideo-p-search" id="subtitle"]] [[!template text="""it, you would have to re-index, you would have to make sure""" start="00:05:27.320" video="mainVideo-p-search" id="subtitle"]] [[!template text="""that file is re-indexed properly. Then you have the""" start="00:05:29.960" video="mainVideo-p-search" id="subtitle"]] [[!template text="""whole issue of, well, if you're searching in Emacs,""" start="00:05:32.560" video="mainVideo-p-search" id="subtitle"]] [[!template text="""you have all these projects, this directory,""" start="00:05:36.120" video="mainVideo-p-search" id="subtitle"]] [[!template text="""that directory, how do you know which? Do you always have to""" start="00:05:38.800" video="mainVideo-p-search" id="subtitle"]] [[!template text="""keep them in sync? It's quite a hard task to handle""" start="00:05:42.480" video="mainVideo-p-search" id="subtitle"]] [[!template text="""that. Then on the other end, tools like ripgrep can""" start="00:05:47.400" video="mainVideo-p-search" id="subtitle"]] [[!template text="""search very fast. Even though they can't search maybe on the""" start="00:05:53.080" video="mainVideo-p-search" id="subtitle"]] [[!template text="""order of tens of thousands of repositories, for a local""" start="00:05:59.120" video="mainVideo-p-search" id="subtitle"]] [[!template text="""setting, they should be plenty fast enough.""" start="00:06:03.920" video="mainVideo-p-search" id="subtitle"]] [[!template text="""I benchmarked. Ripgrep, for example, is""" start="00:06:06.040" video="mainVideo-p-search" id="subtitle"]] [[!template text="""on the order of gigabytes per second.""" start="00:06:12.240" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Definitely, it can search a few pretty big size""" start="00:06:15.960" video="mainVideo-p-search" id="subtitle"]] [[!template text="""repositories.""" start="00:06:19.240" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Next main task. We decided not to use an""" start="00:06:21.757" video="mainVideo-p-search" id="subtitle"]] [[!template text="""index. Next task is how do we rank search results? So there's""" start="00:06:24.800" video="mainVideo-p-search" id="subtitle"]] [[!template text="""two main algorithms that are used these days. The first""" start="00:06:29.960" video="mainVideo-p-search" id="subtitle"]] [[!template text="""one is tf-idf, which stands for term frequency, inverse""" start="00:06:33.440" video="mainVideo-p-search" id="subtitle"]] [[!template text="""target frequency. Then there's BM25, which is sort of a""" start="00:06:36.520" video="mainVideo-p-search" id="subtitle"]] [[!template text="""modified tf-idf algorithm.""" start="00:06:43.040" video="mainVideo-p-search" id="subtitle"]] [[!template text="""tf-idf, without going into""" start="00:06:43.553" video="mainVideo-p-search" id="subtitle"]] [[!template text="""too much detail, essentially multiplies two terms. One""" start="00:06:45.680" video="mainVideo-p-search" id="subtitle"]] [[!template text="""is the term frequency, and then you multiply it by the""" start="00:06:49.160" video="mainVideo-p-search" id="subtitle"]] [[!template text="""inverse document frequency. The term frequency is a""" start="00:06:51.880" video="mainVideo-p-search" id="subtitle"]] [[!template text="""measure of how often that search term occurs. The""" start="00:06:54.560" video="mainVideo-p-search" id="subtitle"]] [[!template text="""inverse document frequency is a measure of how much""" start="00:06:58.520" video="mainVideo-p-search" id="subtitle"]] [[!template text="""information that term provides. If the term occurs a lot,""" start="00:07:00.800" video="mainVideo-p-search" id="subtitle"]] [[!template text="""then it gets a higher score in the term frequency section.""" start="00:07:06.200" video="mainVideo-p-search" id="subtitle"]] [[!template text="""But if it's a common word that exists in a lot of documents,""" start="00:07:08.720" video="mainVideo-p-search" id="subtitle"]] [[!template text="""then its inverse document frequency goes down.""" start="00:07:12.400" video="mainVideo-p-search" id="subtitle"]] [[!template text="""It kind of scores it less. You'll find that words like the,""" start="00:07:13.901" video="mainVideo-p-search" id="subtitle"]] [[!template text="""in, is, these really common words, since they occur""" start="00:07:20.880" video="mainVideo-p-search" id="subtitle"]] [[!template text="""everywhere, their inverse document frequency is""" start="00:07:25.960" video="mainVideo-p-search" id="subtitle"]] [[!template text="""essentially zero. They don't really count towards a""" start="00:07:29.200" video="mainVideo-p-search" id="subtitle"]] [[!template text="""score. But when you have rare words that only occur in a""" start="00:07:32.480" video="mainVideo-p-search" id="subtitle"]] [[!template text="""few documents, they're weighted a lot more. So the more""" start="00:07:35.680" video="mainVideo-p-search" id="subtitle"]] [[!template text="""those rare words occur, they boost the score higher.""" start="00:07:37.680" video="mainVideo-p-search" id="subtitle"]] [[!template text="""BM25 is a modification of this. It's essentially TF, it's""" start="00:07:41.160" video="mainVideo-p-search" id="subtitle"]] [[!template text="""essentially the previous one, except it dampens out terms""" start="00:07:48.840" video="mainVideo-p-search" id="subtitle"]] [[!template text="""that occur more often. Imagine you have a bunch of""" start="00:07:53.120" video="mainVideo-p-search" id="subtitle"]] [[!template text="""documents. One has a term 10 times, one has a term, that same""" start="00:07:55.440" video="mainVideo-p-search" id="subtitle"]] [[!template text="""term a hundred times, another has a thousand times.""" start="00:07:59.360" video="mainVideo-p-search" id="subtitle"]] [[!template text="""You'll see the score dampens off as the number of""" start="00:08:02.440" video="mainVideo-p-search" id="subtitle"]] [[!template text="""occurrences increases. That prevents any one term from""" start="00:08:06.800" video="mainVideo-p-search" id="subtitle"]] [[!template text="""overpowering the score. This is the algorithm I ended up""" start="00:08:10.640" video="mainVideo-p-search" id="subtitle"]] [[!template text="""choosing for my implementation. So with a plan of using a""" start="00:08:16.560" video="mainVideo-p-search" id="subtitle"]] [[!template text="""command line tool like ripgrep to get term occurrences, and""" start="00:08:21.040" video="mainVideo-p-search" id="subtitle"]] [[!template text="""then using a scoring algorithm like BM25 to rank the terms,""" start="00:08:29.560" video="mainVideo-p-search" id="subtitle"]] [[!template text="""we can combine this together and create a simple search""" start="00:08:36.800" video="mainVideo-p-search" id="subtitle"]] [[!template text="""mechanism.""" start="00:08:40.080" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Here we're in the directory for the Emacs source code.""" start="00:08:41.200" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Let's say we want to search for the display code. We""" start="00:08:47.440" video="mainVideo-p-search" id="subtitle"]] [[!template text="""run the p-search command, starting the search engine. It""" start="00:08:53.480" video="mainVideo-p-search" id="subtitle"]] [[!template text="""opens up. We notice it has three sections, the candidate""" start="00:08:58.680" video="mainVideo-p-search" id="subtitle"]] [[!template text="""generators, the priors, and the search results. The""" start="00:09:01.160" video="mainVideo-p-search" id="subtitle"]] [[!template text="""candidate generators generates the search space we're""" start="00:09:05.200" video="mainVideo-p-search" id="subtitle"]] [[!template text="""looking on. These are all composable and you can add as""" start="00:09:10.000" video="mainVideo-p-search" id="subtitle"]] [[!template text="""many as you want. So with this, it specifies that here""" start="00:09:14.720" video="mainVideo-p-search" id="subtitle"]] [[!template text="""we're searching on the file system and we're searching in""" start="00:09:19.720" video="mainVideo-p-search" id="subtitle"]] [[!template text="""this directory. We're using the ripgrep tool to search""" start="00:09:25.240" video="mainVideo-p-search" id="subtitle"]] [[!template text="""with, and we want to make sure that we're searching only on""" start="00:09:30.800" video="mainVideo-p-search" id="subtitle"]] [[!template text="""files committed to Git. Here we see the search results.""" start="00:09:33.360" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Notice here is their final probability. Here, notice""" start="00:09:40.480" video="mainVideo-p-search" id="subtitle"]] [[!template text="""that they're all the same, and they're the same because we""" start="00:09:45.160" video="mainVideo-p-search" id="subtitle"]] [[!template text="""don't have any search criteria specified here. Suppose""" start="00:09:47.080" video="mainVideo-p-search" id="subtitle"]] [[!template text="""we want to search for display-related code. We add a""" start="00:09:50.720" video="mainVideo-p-search" id="subtitle"]] [[!template text="""query: display.""" start="00:09:55.680" video="mainVideo-p-search" id="subtitle"]] [[!template text="""So then it spins off the processes, gets the search term""" start="00:09:57.360" video="mainVideo-p-search" id="subtitle"]] [[!template text="""counts and calculates the new scores. Notice here that""" start="00:10:06.560" video="mainVideo-p-search" id="subtitle"]] [[!template text="""the results that come on top are just at first glance appear""" start="00:10:10.880" video="mainVideo-p-search" id="subtitle"]] [[!template text="""to be relevant to display. Remember, if we compare""" start="00:10:15.760" video="mainVideo-p-search" id="subtitle"]] [[!template text="""that to just running a ripgrep raw, notice here we're""" start="00:10:19.920" video="mainVideo-p-search" id="subtitle"]] [[!template text="""getting 53,000 results and it's pretty hard to go through""" start="00:10:25.080" video="mainVideo-p-search" id="subtitle"]] [[!template text="""these results and make sense of it.""" start="00:10:31.280" video="mainVideo-p-search" id="subtitle"]] [[!template text="""So that's p-search in a nutshell.""" start="00:10:34.320" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Next, I wanted to talk about the story of Flight 447.""" start="00:10:41.457" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Flight 447 going from Rio de Janeiro to Paris""" start="00:10:45.983" video="mainVideo-p-search" id="subtitle"]] [[!template text="""crashed somewhere in the Atlantic Ocean""" start="00:10:49.327" video="mainVideo-p-search" id="subtitle"]] [[!template text="""on June 1st, 2009, killing everyone on board.""" start="00:10:51.510" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Four search attempts were made to find the wreckage.""" start="00:10:54.714" video="mainVideo-p-search" id="subtitle"]] [[!template text="""None of them were successful, except the finding of some debris""" start="00:10:56.895" video="mainVideo-p-search" id="subtitle"]] [[!template text="""and a dead body. It was decided that they really wanted""" start="00:11:01.076" video="mainVideo-p-search" id="subtitle"]] [[!template text="""to find the wreckage to retrieve data as to why the search""" start="00:11:05.480" video="mainVideo-p-search" id="subtitle"]] [[!template text="""occurred. This occurred two years after the""" start="00:11:09.520" video="mainVideo-p-search" id="subtitle"]] [[!template text="""initial crash. With this next search attempt, they""" start="00:11:14.640" video="mainVideo-p-search" id="subtitle"]] [[!template text="""wanted to create a probability distribution of where the""" start="00:11:19.960" video="mainVideo-p-search" id="subtitle"]] [[!template text="""crash could be. The only piece of concrete data they had""" start="00:11:23.200" video="mainVideo-p-search" id="subtitle"]] [[!template text="""was a GPS signal from the ship at 210 containing the GPS""" start="00:11:26.760" video="mainVideo-p-search" id="subtitle"]] [[!template text="""location of the plane was at 2.98 degrees north, 30.59""" start="00:11:35.080" video="mainVideo-p-search" id="subtitle"]] [[!template text="""degrees west. That was the only data they had to go off of.""" start="00:11:40.240" video="mainVideo-p-search" id="subtitle"]] [[!template text="""So they drew a circle around that point""" start="00:11:44.720" video="mainVideo-p-search" id="subtitle"]] [[!template text="""with a radius of 40 nautical miles. They assumed that""" start="00:11:50.080" video="mainVideo-p-search" id="subtitle"]] [[!template text="""anything outside the circle would have been impossible for""" start="00:11:54.680" video="mainVideo-p-search" id="subtitle"]] [[!template text="""the ship to reach. This was the starting point for""" start="00:11:57.480" video="mainVideo-p-search" id="subtitle"]] [[!template text="""creating the probability distribution of where the""" start="00:12:01.240" video="mainVideo-p-search" id="subtitle"]] [[!template text="""wreckage occurred. Anything outside the circle, they""" start="00:12:04.800" video="mainVideo-p-search" id="subtitle"]] [[!template text="""assumed it was impossible to reach.""" start="00:12:08.120" video="mainVideo-p-search" id="subtitle"]] [[!template text="""The only other pieces of data were the four failed search""" start="00:12:09.640" video="mainVideo-p-search" id="subtitle"]] [[!template text="""attempts and then some of the debris found. One thing they""" start="00:12:16.480" video="mainVideo-p-search" id="subtitle"]] [[!template text="""did decide was to look at similar crashes where control was""" start="00:12:21.720" video="mainVideo-p-search" id="subtitle"]] [[!template text="""lost to analyze where the crashes landed, compared to where""" start="00:12:26.160" video="mainVideo-p-search" id="subtitle"]] [[!template text="""the loss of control started. This probability""" start="00:12:30.320" video="mainVideo-p-search" id="subtitle"]] [[!template text="""distribution, the circular normal distribution was""" start="00:12:37.400" video="mainVideo-p-search" id="subtitle"]] [[!template text="""decided upon. Here you can see that the center has a lot""" start="00:12:43.480" video="mainVideo-p-search" id="subtitle"]] [[!template text="""higher chance of finding the wreckage. As you go away""" start="00:12:47.920" video="mainVideo-p-search" id="subtitle"]] [[!template text="""from the center, the probability of finding the wreckage""" start="00:12:51.880" video="mainVideo-p-search" id="subtitle"]] [[!template text="""decreases a lot. The next thing they looked at was, well,""" start="00:12:55.400" video="mainVideo-p-search" id="subtitle"]] [[!template text="""they noticed they had retrieved some dead bodies from the""" start="00:13:02.320" video="mainVideo-p-search" id="subtitle"]] [[!template text="""wreckage. So they thought that they could calculate the""" start="00:13:05.960" video="mainVideo-p-search" id="subtitle"]] [[!template text="""backward drift on that particular day to find where the""" start="00:13:12.960" video="mainVideo-p-search" id="subtitle"]] [[!template text="""crash might've occurred. If they found bodies at a""" start="00:13:18.440" video="mainVideo-p-search" id="subtitle"]] [[!template text="""particular location, they can kind of work backwards from""" start="00:13:21.480" video="mainVideo-p-search" id="subtitle"]] [[!template text="""that in order to find where the initial crash occurred.""" start="00:13:25.120" video="mainVideo-p-search" id="subtitle"]] [[!template text="""So here you can see the probability distribution based off of""" start="00:13:30.666" video="mainVideo-p-search" id="subtitle"]] [[!template text="""the backward drift model. Here you see the darker colors""" start="00:13:34.720" video="mainVideo-p-search" id="subtitle"]] [[!template text="""have a higher probability of finding the location. So""" start="00:13:40.280" video="mainVideo-p-search" id="subtitle"]] [[!template text="""with all these pieces of data, so with that circular 40""" start="00:13:46.160" video="mainVideo-p-search" id="subtitle"]] [[!template text="""nautical mile uniform distribution, with that circular""" start="00:13:50.680" video="mainVideo-p-search" id="subtitle"]] [[!template text="""normal distribution of comparing similar crashes, as well""" start="00:13:54.960" video="mainVideo-p-search" id="subtitle"]] [[!template text="""as with the backward drift, they were able to combine all""" start="00:14:02.200" video="mainVideo-p-search" id="subtitle"]] [[!template text="""three of these pieces""" start="00:14:07.440" video="mainVideo-p-search" id="subtitle"]] [[!template text="""in order to come up with a final prior distribution of where""" start="00:14:08.560" video="mainVideo-p-search" id="subtitle"]] [[!template text="""the wreckage occurred. So this is what the final model""" start="00:14:14.600" video="mainVideo-p-search" id="subtitle"]] [[!template text="""they came upon. Here you can see it has that 40 nautical""" start="00:14:19.520" video="mainVideo-p-search" id="subtitle"]] [[!template text="""mile radius circle. It has that darker center, which""" start="00:14:24.720" video="mainVideo-p-search" id="subtitle"]] [[!template text="""indicates a higher probability because of the""" start="00:14:29.680" video="mainVideo-p-search" id="subtitle"]] [[!template text="""crash similarity. Then here you also see along this line""" start="00:14:32.040" video="mainVideo-p-search" id="subtitle"]] [[!template text="""has a slightly higher probability due to the backward drift""" start="00:14:38.960" video="mainVideo-p-search" id="subtitle"]] [[!template text="""distribution.""" start="00:14:50.800" video="mainVideo-p-search" id="subtitle"]] [[!template text="""So the next thing is, since they had performed searches,""" start="00:14:52.120" video="mainVideo-p-search" id="subtitle"]] [[!template text="""they decided to incorporate the data from those searches""" start="00:14:56.560" video="mainVideo-p-search" id="subtitle"]] [[!template text="""into their new distribution. Here you can see places""" start="00:15:00.560" video="mainVideo-p-search" id="subtitle"]] [[!template text="""where they searched initially. If you think about it,""" start="00:15:04.760" video="mainVideo-p-search" id="subtitle"]] [[!template text="""you can assume that, well, if you search for something,""" start="00:15:08.880" video="mainVideo-p-search" id="subtitle"]] [[!template text="""there's a good chance you'll find it, but not necessarily.""" start="00:15:11.400" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Anywhere where they searched, the probability of it""" start="00:15:14.200" video="mainVideo-p-search" id="subtitle"]] [[!template text="""finding it there is greatly reduced. It's not zero because""" start="00:15:18.440" video="mainVideo-p-search" id="subtitle"]] [[!template text="""obviously you can look for something and miss it, but it kind""" start="00:15:22.840" video="mainVideo-p-search" id="subtitle"]] [[!template text="""of reduces the probability that we would expect to find it in""" start="00:15:26.880" video="mainVideo-p-search" id="subtitle"]] [[!template text="""those already searched locations. This is the""" start="00:15:31.120" video="mainVideo-p-search" id="subtitle"]] [[!template text="""posterior distribution or distribution after counting""" start="00:15:36.680" video="mainVideo-p-search" id="subtitle"]] [[!template text="""observations made.""" start="00:15:41.920" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Here we can see kind of these cutouts of where the""" start="00:15:44.560" video="mainVideo-p-search" id="subtitle"]] [[!template text="""previous searches occurred. This is the final""" start="00:15:48.760" video="mainVideo-p-search" id="subtitle"]] [[!template text="""distribution they went off of to perform the subsequent""" start="00:15:53.960" video="mainVideo-p-search" id="subtitle"]] [[!template text="""search. In the end, the wreckage was found at a point close to""" start="00:15:57.000" video="mainVideo-p-search" id="subtitle"]] [[!template text="""the center here, thus validating this methodology.""" start="00:16:02.000" video="mainVideo-p-search" id="subtitle"]] [[!template text="""We can see the power of this Bayesian search methodology""" start="00:16:06.771" video="mainVideo-p-search" id="subtitle"]] [[!template text="""in the way that we could take information from all the sources we had.""" start="00:16:10.333" video="mainVideo-p-search" id="subtitle"]] [[!template text="""We could draw analogies to similar situations.""" start="00:16:14.000" video="mainVideo-p-search" id="subtitle"]] [[!template text="""We can quantify these, combine them into a model,""" start="00:16:19.238" video="mainVideo-p-search" id="subtitle"]] [[!template text="""and then also update our model according to each observation we make.""" start="00:16:22.480" video="mainVideo-p-search" id="subtitle"]] [[!template text="""I think there's a lot of similarities to be drawn with""" start="00:16:27.894" video="mainVideo-p-search" id="subtitle"]] [[!template text="""searching on a computer in the sense that when we search for""" start="00:16:30.360" video="mainVideo-p-search" id="subtitle"]] [[!template text="""something, there's oftentimes a story we kind of have as to""" start="00:16:35.160" video="mainVideo-p-search" id="subtitle"]] [[!template text="""what search terms exist, where we expect to find the file.""" start="00:16:39.400" video="mainVideo-p-search" id="subtitle"]] [[!template text="""For example, if you're implementing a new feature, you'll""" start="00:16:43.960" video="mainVideo-p-search" id="subtitle"]] [[!template text="""often have some search terms in mind that you think will be""" start="00:16:46.720" video="mainVideo-p-search" id="subtitle"]] [[!template text="""relevant. Some search terms, you might think they have a""" start="00:16:49.920" video="mainVideo-p-search" id="subtitle"]] [[!template text="""possibility of being relevant, but maybe you're not sure.""" start="00:16:54.720" video="mainVideo-p-search" id="subtitle"]] [[!template text="""There's some directories where you know that they're not""" start="00:16:57.600" video="mainVideo-p-search" id="subtitle"]] [[!template text="""relevant. There's other criteria like, well, you know that""" start="00:17:02.880" video="mainVideo-p-search" id="subtitle"]] [[!template text="""maybe somebody in particular worked on this code.""" start="00:17:07.760" video="mainVideo-p-search" id="subtitle"]] [[!template text="""What if you could incorporate that information? Like, I know""" start="00:17:11.400" video="mainVideo-p-search" id="subtitle"]] [[!template text="""this author, he's always working on this feature. What if""" start="00:17:16.320" video="mainVideo-p-search" id="subtitle"]] [[!template text="""I just give the files that this person works on a higher""" start="00:17:21.400" video="mainVideo-p-search" id="subtitle"]] [[!template text="""probability than ones he doesn't work on? Or maybe you think""" start="00:17:25.520" video="mainVideo-p-search" id="subtitle"]] [[!template text="""that this is a file that's committed too often. You think""" start="00:17:32.600" video="mainVideo-p-search" id="subtitle"]] [[!template text="""that maybe the amount of times of commits it receives""" start="00:17:38.600" video="mainVideo-p-search" id="subtitle"]] [[!template text="""should change your probability of this file being""" start="00:17:43.440" video="mainVideo-p-search" id="subtitle"]] [[!template text="""relevant. That's where p-search comes in.""" start="00:17:47.720" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Its aim is to be a framework in order to incorporate all these""" start="00:17:52.840" video="mainVideo-p-search" id="subtitle"]] [[!template text="""sorts of different prior information into your searching""" start="00:17:57.680" video="mainVideo-p-search" id="subtitle"]] [[!template text="""process. You're able to say things like, I want files""" start="00:18:01.360" video="mainVideo-p-search" id="subtitle"]] [[!template text="""authored by this user to be given higher probability. I want""" start="00:18:06.000" video="mainVideo-p-search" id="subtitle"]] [[!template text="""this author to be given a lower priority. I know this author""" start="00:18:11.120" video="mainVideo-p-search" id="subtitle"]] [[!template text="""never works on this code. If he has a commit, then lower its""" start="00:18:13.920" video="mainVideo-p-search" id="subtitle"]] [[!template text="""probability, or you can specify specific paths, or you can""" start="00:18:18.760" video="mainVideo-p-search" id="subtitle"]] [[!template text="""specify multiple search terms, weighing different ones""" start="00:18:24.680" video="mainVideo-p-search" id="subtitle"]] [[!template text="""according to how you think those terms should be relevant.""" start="00:18:30.200" video="mainVideo-p-search" id="subtitle"]] [[!template text="""So with p-search, we're able to incorporate information""" start="00:18:38.920" video="mainVideo-p-search" id="subtitle"]] [[!template text="""from multiple sources. Here, for example, we have a prior""" start="00:18:42.080" video="mainVideo-p-search" id="subtitle"]] [[!template text="""of type git author, and we're looking for all of the files""" start="00:18:46.280" video="mainVideo-p-search" id="subtitle"]] [[!template text="""that are committed to by Lars. So the more commits he has,""" start="00:18:52.080" video="mainVideo-p-search" id="subtitle"]] [[!template text="""the higher probability is given to that file. Suppose""" start="00:18:56.720" video="mainVideo-p-search" id="subtitle"]] [[!template text="""there's a feature I know he worked on, but I don't know the""" start="00:19:01.400" video="mainVideo-p-search" id="subtitle"]] [[!template text="""file or necessarily even key terms of it. Well, with this, I""" start="00:19:04.560" video="mainVideo-p-search" id="subtitle"]] [[!template text="""can incorporate that information.""" start="00:19:09.160" video="mainVideo-p-search" id="subtitle"]] [[!template text="""So let's search again. Let's add display.""" start="00:19:12.141" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Let's see what responses we get back here. We can add""" start="00:19:16.000" video="mainVideo-p-search" id="subtitle"]] [[!template text="""as many of these criteria as we want. We can even specify that""" start="00:19:22.960" video="mainVideo-p-search" id="subtitle"]] [[!template text="""the title of the file name should be a certain type. Let's""" start="00:19:27.200" video="mainVideo-p-search" id="subtitle"]] [[!template text="""say we're only concerned about C files. We add the file""" start="00:19:31.520" video="mainVideo-p-search" id="subtitle"]] [[!template text="""name should contain .c in it. With this, now we""" start="00:19:36.600" video="mainVideo-p-search" id="subtitle"]] [[!template text="""notice that all of the C files containing display authored""" start="00:19:45.400" video="mainVideo-p-search" id="subtitle"]] [[!template text="""by Lars should be given higher probability. We can""" start="00:19:51.320" video="mainVideo-p-search" id="subtitle"]] [[!template text="""continue to add these priors as we feel fit. The workflow""" start="00:19:56.280" video="mainVideo-p-search" id="subtitle"]] [[!template text="""that I found helps when searching is that you'll add""" start="00:20:02.720" video="mainVideo-p-search" id="subtitle"]] [[!template text="""criteria, you'll see some good results come up and some bad""" start="00:20:07.520" video="mainVideo-p-search" id="subtitle"]] [[!template text="""results come up. So you'll often find a pattern in those""" start="00:20:11.360" video="mainVideo-p-search" id="subtitle"]] [[!template text="""bad results, like, oh, I don't want test files, or this""" start="00:20:15.320" video="mainVideo-p-search" id="subtitle"]] [[!template text="""directory isn't relevant, or something like that. Then""" start="00:20:18.840" video="mainVideo-p-search" id="subtitle"]] [[!template text="""you can update your prior distribution, adding its""" start="00:20:22.680" video="mainVideo-p-search" id="subtitle"]] [[!template text="""criteria, and then rerun it, and then it will get different""" start="00:20:27.200" video="mainVideo-p-search" id="subtitle"]] [[!template text="""probabilities for the files. So in the end, you'll have a""" start="00:20:31.120" video="mainVideo-p-search" id="subtitle"]] [[!template text="""list of results that's tailor-made to the thing you're""" start="00:20:35.160" video="mainVideo-p-search" id="subtitle"]] [[!template text="""searching for.""" start="00:20:37.640" video="mainVideo-p-search" id="subtitle"]] [[!template text="""There's a couple of other features I""" start="00:20:40.405" video="mainVideo-p-search" id="subtitle"]] [[!template text="""want to go through. One thing is that each of these priors,""" start="00:20:41.640" video="mainVideo-p-search" id="subtitle"]] [[!template text="""you can specify the importance. In other words, how""" start="00:20:49.080" video="mainVideo-p-search" id="subtitle"]] [[!template text="""important is this particular piece of information to your""" start="00:20:55.840" video="mainVideo-p-search" id="subtitle"]] [[!template text="""search? So here, everything is of importance medium. But""" start="00:21:01.120" video="mainVideo-p-search" id="subtitle"]] [[!template text="""let's say I really care about something having the word""" start="00:21:05.200" video="mainVideo-p-search" id="subtitle"]] [[!template text="""display in it. I'm going to change its importance.""" start="00:21:07.880" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Instead of medium, I'll change its importance to high.""" start="00:21:12.680" video="mainVideo-p-search" id="subtitle"]] [[!template text="""What that does essentially is things that don't have""" start="00:21:18.600" video="mainVideo-p-search" id="subtitle"]] [[!template text="""display in it are given a much bigger penalty and things with""" start="00:21:23.280" video="mainVideo-p-search" id="subtitle"]] [[!template text="""the word display in it are rated much higher.""" start="00:21:28.080" video="mainVideo-p-search" id="subtitle"]] [[!template text="""With this, we're able to fine-tune the results that we get.""" start="00:21:28.129" video="mainVideo-p-search" id="subtitle"]] [[!template text="""Another thing you can do is that you can add the complement or""" start="00:21:38.560" video="mainVideo-p-search" id="subtitle"]] [[!template text="""the inverse of certain queries. Let's say you want to""" start="00:21:45.640" video="mainVideo-p-search" id="subtitle"]] [[!template text="""search for display, but you don't want it to contain the word""" start="00:21:49.760" video="mainVideo-p-search" id="subtitle"]] [[!template text="""frame. With the complement option on, when we create this""" start="00:21:53.240" video="mainVideo-p-search" id="subtitle"]] [[!template text="""search prior, now it's going to be searching for frame, but""" start="00:21:58.040" video="mainVideo-p-search" id="subtitle"]] [[!template text="""instead of increasing the search score, it's going to""" start="00:22:01.840" video="mainVideo-p-search" id="subtitle"]] [[!template text="""decrease it if it contains the word frame.""" start="00:22:04.960" video="mainVideo-p-search" id="subtitle"]] [[!template text="""So here, things related to frame are kind of""" start="00:22:07.000" video="mainVideo-p-search" id="subtitle"]] [[!template text="""deprioritized. We can also say that we really don't want""" start="00:22:14.320" video="mainVideo-p-search" id="subtitle"]] [[!template text="""the search to contain the word frame by increasing its""" start="00:22:18.080" video="mainVideo-p-search" id="subtitle"]] [[!template text="""importance. So with all these composable pieces, we can""" start="00:22:21.600" video="mainVideo-p-search" id="subtitle"]] [[!template text="""create kind of a search that's tailor-made to our needs.""" start="00:22:27.200" video="mainVideo-p-search" id="subtitle"]] [[!template text="""That concludes this talk. There's a lot more I could talk""" start="00:22:33.413" video="mainVideo-p-search" id="subtitle"]] [[!template text="""about with regards to research, so definitely follow the""" start="00:22:35.760" video="mainVideo-p-search" id="subtitle"]] [[!template text="""project if you're interested. Thanks for watching, and I""" start="00:22:37.800" video="mainVideo-p-search" id="subtitle"]] [[!template text="""hope you enjoy the rest of the conference.""" start="00:22:40.640" video="mainVideo-p-search" id="subtitle"]]