LangChain Version 0.1 Explained | New Features & Changes

Key Takeaways

good morning everyone how is it going today welcome to this video and welcome back to the channel it has been a while now in this video we're going to be talking about Lang chain and specifically the new version of Lang chain which is its first stable release and we're going to be talking we're going to be going through the official article that they posted in January 8th so a few days ago about this new release we're going to be talking about the main changes that they have made to the library and what you will have to do if you're planning on building a new application using this new release of the framework okay now this one has been an amazing year here at the company we have been able to develop several applications using Lang chain and these applications are already working they're already in production and they are already solving real world problems and helping real people with their lives so I'm very happy to see that Lang chain is already getting to its first stable release um so yeah without any further Ado uh let's go right to the video and don't forget that if you like content about artificial intelligence and uh software development more specifically tutorials on how to implement them don't forget to subscribe okay so let's get right into [Music] it all right so let's start with the article right here um this is the article that was published when the official release was announced it was in January 8th so it's just a few days ago and we're going to go through all of what the article discusses and to explain the new features and changes right here okay now Lang chain as you may know is a library that is released both in Python and JavaScript um that makes it available for your backend applications here you have the links to the guides to the documentation and you also have here the link to the python GitHub discussion and the link to the YouTube walkthr made by Harrison himself who is the CEO and the creator of Lang chain okay I will leave of course the link to this article in in the description so that you can check it out um about the introduction we're we're presented with a structural change which is the main and the most important change that the that Lang is going through um the first one is that instead of considering it a library now we should start considering Lang chain as a framework a framework for developing uh language model applications okay um what does this mean this means that the Library part in its in itself is being split into different packages okay and that's exemplified right here before if you wanted to use Lang chain what you would do is you would install the lanching package if you were using for example python you would do pip install lanching and then you would have everything inside of that package from there you would you were able to install all of the Integrations with the third party providers like for example the vector stores you would do from Lang chain. Vector stores import chroma or quadrant for example okay now what happens is that they have split Lang chain itself into three different packages they have what is Lang chain right here they have what is Lang chain Community right here and then they have Lang chain core okay it's easier to see in this way here right here so what is Lang chain core Lang chain core is going to be basically where you will find all of the basic abstractions of Lang Ching okay here you will find for example The Prompt classes the AI message classes the llm classes all of these abstractions that allow you to um use sort of building blocks to create your applications okay that's Lang chain core okay now secondly we have Lang chain Community which is going to be the the in the third party Integrations of langing okay here in the article they say that they have almost 700 Integrations ranging from language models to Vector stores to tools for the agents of Lang Ching to use um in and all of these Integrations or almost all of them are going to be included in the langing community package so if you do pip install langing core you will not have the packages that pertain to the uh third party Integrations okay so for example now if you want to use chroma as a vector store you will not be able to import it from langin itself but you will have to do you will have to to do pip install langin community and then from langin community. Vector stores import chroma for example I mean I don't know exactly the path from the top of my head but that's that's kind of how it's going to look like all right um so that's the the longchain community right now you have also the model providers the retrievals and the tools like for example the ones that allow you to to serve the web uh Etc okay so that's the second part right here um however something to keep in mind is that not all of the third-party Integrations are going to be inside of the langing community package um so far as they say they have 10 packages that are outside from the langin community package and they have their own package among this we can find open AI Google and mistel for example okay that means that they have their own package that you will have to install if you are going to if you're going to to use them so for example if you want to use a language model by open AI you will have to pep install Lang chain open AI uh and then that's the one that is going to include uh the language models and the embeddings from open AI okay so that's for third party Integrations and thirdly we have the Lang chain The Irregular Lang chain package that one right there they have left it in the version 0.0 point x um because of course they don't want to break any already functioning applications so yeah um so yeah the main features of this release are going to be located in langin community and langin core and that's um from where you should start importing your your components in order to use the latest version okay all right and we arrived to the part of observability um as you may know language model applications are very non-deterministic and I mean there is a non-deterministic component to them in a sense that we are never absolutely sure of what a language model is going to respond to our input that is why observability is a main part of um building language model applications and that is actually exemplified right here here the observability is going to surround all of the framew work um in order to solve this problem of non non-deterministic components uh Lang chain they have built a tool called lsmith which is basically a tool that is going to allow you to log every single input and output for every element inside of your chain so that you have the best I mean a better debugging experience of your application uh I haven't tried it yet they currently they have it in a private beta version but if you want to check it out you just have to go and click right here in the lsmith Pod and you can just sign up to their weight list um so yeah that's basically what they're building and it's more of a platform than than the library or anything else it's going to be um a place where you're going to be able to debug your language model applications um as far as I know this is the first uh product that does this in the market but I mean there might be others now we also have right here an accent or a mention of composability um now what is composability it basically means that you're supposed to be able to chain different methods and components together which I suppose that is very useful to you if you're using Lang chain because as the name suggests it is a library or a framework that allows you to chain different different logic Parts logic components together to create an application or a chain okay um so we were already building chains and this time or I mean a few months ago already I think they they released what they call Lang chain expression language which is right here and they call it LC this language expression this langin expression language basically is just a way to write and to declare your your chains in a more pipeline like um syntax okay I think that we can see it right here an example um how to create um chain so yeah for example you would just create your different components of your chain and then you put them all together in a chain like this and this is supposed to to work for all of the chains available in Lang chain this is of course going this all of course makes it possible for you to modify already existing chains and then also some changes in the naming conventions but all right we we will see that when we when we're actually using the library all right something something else that they mentioned right here is their emphasis on the streaming end points okay that they are going to be exposing in the langu in the Lang chain um in the Lang chain expression language okay um now what is streaming in in language model applications as you may have noticed in chat gbt you are not getting the entire response and the entire generated response in one batch okay so let's say that in chpt you ask a question and the answer is this big you're not going to get all of this in one batch if you have noticed when you ask a question it starts um showing the answer as it is being generated it's it's like it's the as if the language models was right in front of you okay now this is important not not only um to make it look good but for the user experience in general otherwise the users would have to wait like several seconds before they get an answer so showing them in real time that the their answer is being generated is very important for the for the experience in our language model applications so that's why Lang chain they are putting out this their exposure using two different methods stream and a stream inside of any chain constructed with language uh with L chain expression language so now it's going to be even easier for us to stream our answers into into our applications and make it easier for our users okay uh we're getting to uh the section talking about output parsing which is basically the these are basically functions that allow you to be sure that a certain step in your chain is always going to return the same format as an answer okay so for example if for one step in your chain you need to get an answer in a string format there is an output parser that is I think it's called string output parser that is going to make sure that every time that you get a response from a given um from a given language model or whichever step you're in your chain uh you're going to get a string in return so this is basically uh something that they they have I mean they they already had in Lang chain but this time they are mentioning that for for this output parsers we're going to also be able to stream this output the the content that is being parsed through the output as it is being generated so that is that's also very very useful it's going to make it faster for our chains to work and easier for a developer experience in general okay all right now we arve to the section called retrieval um if you had seen some of the videos in the in this channel you know that one of the main aspects of Lang chain is being able to talk to your personal data okay um being able to ask a language model about data that it was not trained in data that you may have in your in your email account in your nodes in your personal private database or in your company's database and that is one of the main uses of Lang chain um they just reinstantiate that and they also included they they mentioned here that they are in including new Advanced retrieval strategies from Academia you may want to take a look at them if you're interested in the more um technical side of this but something very interesting right here is that they are exposing and indexing API that you can see right here so the indexing API uh we're going to make a video about it and show you how it works and everything but just so that you know what's new about it is that it's um it is here to make you to make it easier to index your data basically so this includes avoid writing duplicated content into your vector store I know that's usually um a delicate part of the application development process in which we have to make sure that every Vector has its its very precise metadata to be sure that we're not duplicating content so this right this one right here I mean this API is here to help us do that also it is supposed to avoid rewriting onchanged content and avoid recomputing embeddings over unchanged content so basically a more advanced indexing for our retrieved I mean for for our data that we're going to use for retrieval um and yeah they just mentioned some other other libraries that use Lang chain for a more opinionated approach and retrieval and here we have a section mentioning the agents now the agents are probably the main or the most popular part of Lang chain which and probably the reason why they become so popular why Lang chain became so popular in the first place um now what are agents in case you're not familiar with it it is a component in your language model application that is able to think by itself and Implement a solution based on what it reconed okay so it will take an input it is going to use the a language model to reason about what what it can do to implement a solution given your question and then it's going to be able to use tools to implement that solution um so for example a tool can be um an external API it can be a calculator it can be for example an API that serves the web and finds and finds website relevant websites relevant to the question that the user asked um so yeah I mean basically it's a a tool that's going to be able to reason and apply a tool based on what youve thought about it um here in the article they say that they have um reinvested a lot of time um covering Integrations with more thirdparty tools in the existing agents that come with lanching um they have also um worked on ways to structure language model responses to fit the input schema of those tools and a more flexible way to specify the the ways in which the agent um the agent tool tools work together and this is what we already mentioned which is the Lang chain expression language okay which is the one that allows you to create chains or agents in a more pipeline uh like synex um they also mentioned that they have implemented new methods agent methods from Academia one of them for example is react which you may see if you go right here they have the official I mean they have the paper if you want to take a look at it um the more technical side but yeah here basically they covered agents which are one of the main parts of flank chain and then finally we just see um I mentioned that they are already thinking of lunching 0.2 which of course is going to be um a a minor release um oh something that I forgot to mention probably is that with this new release they are also in changing the version in convention in a way that every minor bomb any any minor new version is going to be accompanied by a bump in the second digit and these ones are going to have breaking changes and any Buck fixes are going to come with a bump in the third digit um yeah so basically that's pretty much all for this article and um I mean they also have a mention of langra which is another product that they're working on but we will be able to talk about it in another video but so far this is how it looks like this is the what Lang chain is uh promising for us and yeah I mean to me it sounds pretty good let me know what you think about it and be sure to to come here for the next video because we're going to be making the quick start and Showcase of almost the main change es with the code itself um in a Jupiter notebook so that you can get your hands diry with some code So yeah thank you very much for for watching and I'll see you in the next one