Movie Recommender System in Python with LLMs

Key Takeaways

what is going on guys welcome back in this video today we're going to build a movie recommender system using a large language model and a vector store in Python so let us get right into it not [Music] AED all right so we're going to build a movie recommender system in Python today which is going to be powered by a large language model which runs locally on our machine now I'm going to give you a brief sketch of the process so that you understand how the reom Commendation process works and you don't have to just copy paste the code without understanding what it does so it's not going to be the most beautiful sketch here but I'm going to try to explain it as simply as possible what we're going to do is we're going to work with a data set from kaggle and this is going to be a Netflix data set containing information about TV shows and movies and it has features like title of uh the movie or show and description and cast and release year and genre and so on so each row represents a movie or a TV show and has features like the type is it a movie or TV show uh stuff like the title and the description and the release here and the director and the cast and so on so we have a couple of features here per row and what we want to do now is we want to take these and turn them into a textural representation now this is quite simple we don't have to do anything else but just create a string that says something like type is whatever for example movie then title is whatever the title of the movie is and so on and all of this is going to be one uh one large string containing all the information about that movie now what our large language model will do with that string is it will take this string and it will embed it into Vector space to keep it simple this basically means that the large language model does something with this content with this string and turns it into a vector into a high dimensional vector to be precise we're going to use llama 2 in this video uh locally here and llama 2 will produce a vector for each row with the size 4,096 so we're going to have 4,096 different values here numbers that will position this Vector in the vector space in the 4,096 dimensional Vector space this is what we're going to do here so every movie will end up being some point in this Vector space some Vector in this vector space now the idea is that similar movies or movies that are somehow uh yeah similar to one another will end up closer in the vector space than movies that are very different now how does it happen that's the intelligent of that's the intelligence of Lama 2 so we don't have to do any fancy coding here we don't have to implement some intelligent algorithm we don't have to train a machine learning model we expect llama 2 to have the intelligence this is already inside of the model the intelligence needed to do that to take movies that are somewhat similar and to put them together closer together in the vector space than movies that are completely different that's the basic idea and then what we do is we store all of these vectors into a vector store so a vector database we're going to use face which is a uh Facebook I don't know what the acronym is but it's a vector store by Facebook and this Vector store will contain all the different vectors and what we can do then with this Vector store is we get some new movie with the string title uh and type and director and so on as a string we feed this into the index into our Vector store and as a result we get for example the top five most similar movies that could be interesting to you if you like this one so you get a list of five movies that could be interesting to you that is what we're going to build in this video today using llama 2 running on our system locally if you can do that otherwise you will probably have to um to host it somewhere or you have to use the API of something else you can also use chat GPT but you have to pay for it or actually just GPT not chat GPT but if you have uh the resources you can just run this locally using olama so this is what we're going to build in this video today now how are we going to do that we're going to start by first of all installing olama now olama is quite easy to install you just go to the website olama com you download it for your op operating system on Linux you just run this command and then you can use it on your system if you have troubles with that I have a video on this channel about olama where you can look uh where you can see how I install it and and how it works basically but it's really not complicated what we're going to do then is once you have ol Lama installed you're going to type into your terminal ol Lama pull and then the model that you want to use I as I said I'm going to use llama 2 you can also provide another uh model name I think on uh their website you should be able to see the models that they offer you have llama 3 you have mistol you have llama 2 with the different parameters also as well if you want to have a a larger model but you can do that uh and install the model that you want to use all right so once you have that done we're going to also download a data set from kaggle the Netflix movies and TV shows data set you will find a link in the description down below just download the Netflix title CSV v file and then we can get started with the coding now actually before we get started with the coding we also need to install the python packages that we're going to use in this video and for this we're going to open up a command line and install numpy pandas face and requests these are the four packages that we're going to need in this video today numpy and pandas obviously because we're going to work with data face because that is our Vector store and requests because we need to send requests to olama uh now now actually we cannot say face I think we need to say face CPU or face GPU so depending on whether you have a GPU that you can use here run face GPU or face CPU if you want to use your processor for this just install the packages and once you have them we can get started all right so the first thing is we're going to obviously import pandas aspd and we're going to then say the data frame is going to be PD read CSV Netflix title CSV and then we can look at it and you can see we have let me just close this here uh we have a couple of features the IDE of the show we have the type of the show movie or uh TV show we have the title we have the director the cast the country that it was produced in I guess uh the date it was added to Netflix the release year also some rating uh the duration and also the genres here so what kind of movie is it and the most important thing I assume is going to be the description because it tells us briefly what the movie is about and I think that's the most important thing uh maybe together with the documentaries and with a cast that is going to be relevant for uh a similarity search all right so we're going to keep it simple here all we're going to do is we're going to craft for each line a string that represents this or the individual movies uh textually so we're going to create a function down here we're going to call it def uh we're going to call it create textual representation given a row from the data frame and what we're going to do is we're going to say textual representation is going to be equal to a multi-line string like this um and the important thing is to not have any tabs here and we're going to say something like type and type is going to be just so uh actually we need to make this a formatted multi-line string like this the type is going to be the type then we're going to say that the title is going to be the title and actually we can then just go and copy paste this I think and we can say we want to have the director we want to have the cast and we want to have uh the release year and genre and the most important thing also in the end the description so this is going to be description this is going to be actually we need to turn this into row title and into row director and into row cast and into row release I think release year right and into row and I think it was what was it called listed in is what we're looking for listed in and then finally of course row description so that should work and this function now applied to the rows and we're going to do it like this here this function applied of course we need to return the textual [Music] representation this function now applied to the individual roles will give us textual representation uh representations for the individual R so I can say DF apply and then I can just apply the create textual representation function axis equals 1 and I will get for every single um row here I will get the string so I can say actually that this is my textual representation column then you can see I have it and then I can just get the [Music] column and show the different values so if I print the first one we're going to see that this is what what this looks like in the end so this is just our basic representation now which we're going to use to ask llama 2 to turn this into a vector into Vector space or in the vector space so what we're going to do next is we're going to say import Face our Vector store import requests so that we can send requests to llama 2 or to O Lama and import numpy SNP we're going to need this here to create uh to create the array for our uh Vector store we're going to define the dimension of our output as I set to be 496 because that is what is going to be returned by llama 2 uh for the embedding Dimensions we're going to say index is equal to face index flat L2 with a dimension here as a parameter this is basically our database you could say our Vector store that we're creating and then we're going to initialize an X full of zeros and the dimensions here are going to be the length of the textural representations so how many instances do we have how many movies or shows do we have and how large is the dimension for each of them so we're going to have n vectors here this amount of vectors uh off size dimension and the data type is going to be float 32 all right so that is now an array full of zeros as you can see and this is going to be filled up with the embeddings from llama 2 so we're going to say now 4 I uh actually 4 I Row in or actually for I representation in enumerate data frame textual [Music] representation what we want to do is want to say if I is 200s or not 200 if I is divisible by 200 so every 200 uh row I want to print the progress so I know how long it takes I'm not going to run all of this on camera because it's going to take quite some time but you can run this here to see the progress just say uh something like processed and then I and then processed I Str I uh instances for example and uh then we're going to say that I want to send a request which is going to return a response so rest is equal to request I want to send a post request to ol Lama and ol Lama by default unless you change that is running at HTTP and then Local Host Port um 11434 so 11434 is the port that ama is running at and we want to use the embeddings API so/ API SL embeddings and what we want to send to this API is a Json object and this Json object will will contain the model that we want to use so the model is going to be llama 2 and also the prompt that we want to use uh or the prompt that should be embedded so the prompt is going to be equal to the representation that we're currently at there you go and this needs to be closed like this actually this should be indented uh yeah let's let's just leave it like this and this is going to return a response and this response will have a field embedding so we want to say embedding is equal to the response get me the Json object from the response and what I'm interested in is the field embedding this embedding is the vector so all we have to do now is we have to go to the I position this is why we enumerate this uh we going want to go to the I position in our zero vector and we want to replace the zeros with this new embedding that we got from the large language model so NP array embedding and then finally in the end we can add all of x to our Vector store now I will turn this number down to let's say 30 so that you can see how fast this works we have remember uh 8, 87 rows to be processed and if I run this you can see that it isn't that fast uh actually now I need to check that this is equal to zero and then I can run this again process zero instances then it takes some time process 30 instances and so on I think my video is lagging I'm pretty sure my video was lagging while this was running because it was running on the GPU so I'm not going to run this here on camera what I have actually done is I have trained this already and I will show you how you can save this when you've trained it already so you run this you wait for it to finish and once it's finished what you do is you say face right index and then you just take the index so this object here that you created and you write it to a file for example let's call it index as well once you do that you're going to create a file index that contains the index so you only have to run this once then it's done and then you can use it because you now have a vector database full of embeddings for the different movies um and in order to load it then which is what I'm going to do now in order to load it again you just need to go and say um you just need to go and say index is equal to pH read index and then the file for for example index and this is what I'm going to do here now so now I just copy pasted the exported Vector store the index file from my prepared code and all I have to do now is I have to open up a new cell and say that the index is equal to phase re Del index and then I just have to provide index and then I can go ahead and use this index to do a similarity search so how exactly can I now recommend a movie based on another movie now first of all I can craft my own textural representation of whatever movie I like I don't have to use a movie that's already part of the data frame but let's start with a movie that's part of the data frame so what I can do for example is I can say one of my favorite movies is Shutter Island I want to see if that's part of the data frame so let's go and say DF where DF title string contains and let's see if we have Shutter Island in here and in this case we have the movie Shutter and we have the movie Shutter Island so let's copy this ID and let's say that what I want to use now here as a base uh for my recommendation is my favorite movie is actually DF iock and then this so we can see that my favor movie is shudder Island here and now I can just go ahead and I can get the or actually let's go and uh write the code directly what we need to do now is we need to send this movie to O Lama to be embedded now we're going to pretend that this movie is not already embedded so we're not going to use it directly we're going to actually embed it again because this is now our new movie that we passed to the uh to the recommender system here and we want it to be embedded into the vector space and then it should find the most similar movie so the closest embeddings to the embeddings uh or to the embedding of this particular movie so what we do is we say response equals requests. poost and we post to the same URL so Local Host uh Port 11434 SL API embeddings and then we want to say that the Json is going to be equal to a dictionary and the content here is going to be model again llama 2 and the prompt in this case is going to [Music] be our favorite movie and from it the textual representation like this that is going to be our uh requested we sent here and this is going to result in a embedding so in an embedding so we're going to get the embedding from this response again get the Json object get the embedding and uh actually we need to do some additional processing so we're going to say here uh we're going to turn this into a numpy array that's just some formatting stuff here so we're going to say it's a numpy rate off a list of the embedding with a data type float 32 and this is now what we're going to use as a basis for the similarity search so we're going to search our index for similar embeddings to this one so we're going to say d and I is equal to index search given the embedding give me the top five matches so I can say five here and this will result in the top five matches now what I'm going to get here as a response is of course not the actual title not the actual data not the textual representation I will get the indices I will get the actual um we can look at it um I will get the numbers I will not get any specific uh movies so what I have to do is I have to take the data frame and get these particular movies so what I can do is I can say the best matches are going to be equal to NP array and then data frame textual [Music] representation um and from this we're going to go and get I flatten get all these indices flatten them and use them as an index here and then I can just say for match in best matches uh I want to print next movie then I want to print the match so the movie representation and then just an empty line and this is going to give me the top five recommendations now obviously the most similar um match here to my prompt to my embedding is the embedding itself because Shutter Island was already part of the vector store so of course the closest match is shter Island itself the second closest one is this one Devil's gate uh seeking a missing woman in North Dakota an FBI agent and a sheriff focus on her religious zot husband but discovers something far more Sinister sounds like something mysterious and something slightly scary which is also a little bit like shutter Island I've not watched this movie but I think that could be a good match at least from the description it sounds like that um and then you can also see the other matches here now the interesting thing is I don't have to provide actually an existing uh movie I can also make up a movie I can maybe say I can imagine a movie so let's go and copy this and let's turn this into a string and let's just make up something so let's say my title my movie title is the mysterious python the director is who did chter Island let's use the same director and then let's say the cast is Leonardo decapo and then maybe I don't know uh I don't know let let's go with someone else like uh Sylvester is it written like this alone I hope so I hope this is not too embarrassing I'm not a movie guy so I don't know uh let's just say these are the two people and then it's released 2020 and it's mystery drama thriller or something like this and then we can say that the description is uh a group of adventurers discover a mysterious programming snake in the jungle and uh I don't know and find something extremely shocking yeah let's say this is my movie description now I can take this and I can actually save this now as representation and then I can use that as an input here representation then I get the embeddings and the best matches are the following a patch of ful this is when a guard catches a rider television host shoplifting instead of turning him in he only asks to be a friend then begins to rule his life okay I don't know why that is super similar to what I said uh but yeah I mean you can play around with that and see how good the recommendations are maybe it's also reasonable to only include the description destion or to only include genres in description something like this maybe the other things are too um too confusing but that's the basic idea this is how you can build a recommender system based on any data set you just pick a textual representation you uh or you create a textual representation you embed these representations and then you just perform a similarity search and you hope that the embeddings are somewhat intelligent based on the large language model that you use to create them so that's it for today's video I hope you enjoyed it and I hope you learned something if so let me know by hitting a like button and leaving a comment in the comment section down below and of course don't forget to subscribe to this Channel and hit the notification Bell to not miss a single future video for free other than that thank you much for watching see you in the next video and bye