Is Langchain BAD for AI App Devs?

Key Takeaways

line chain has been one of the most popular python libraries to build llm apps but recently it's been in a big soup with a lot of people saying that Lang chain is over bloated and in this video we are going to explore what is the problem with Lang chain and what is the situation in which you should ideally use Lang chain and maybe you should not use lamp chain lamp chain like to be honest like Langston is an amazing Library if you want to get started with llm applications if you want to build AI applications but then there are certain places where the launch and Fanboys overuse Lang chain when they shouldn't have already used launching so we start with what is the problem with land chain so the problem was highlighted by this Reddit user link chain is pointless as the post of the Reddit and starts with the being a rapper so embeddings in launching is a do nothing wrapper for sentence Transformers and its sentence Transformers is an open source library and if you want to use embedding that is a problem the problem is also with the Lang chains readme which is like a grandiose and vague and one of the biggest problems that they have highlighted in this particular user is that Lang chain is highly modularized and it reports a lot of things when you import the data type is not standard and with that if we move on to the next issue that that has been pointed out by mini maxer who is one of the most popular data science voices I would say but any maxer wanted to write a blog post called the problem with the land chain but did not do it because Minimax didn't want to be that guy who criticizes a open source project in good faith but leaving that aside leaving the criticism aside and if you want to look at Lang chain with just a neutral point of view you can see the points that Minimax has raised is Lang chain encourages to lock in for little developer benefit as noted by the original poster the Reddit post there is no inherent advantage in using them and some have some optimal implementations the current implementations of react workflow and the prompt engineering are based on instant GPT which is like text dobbyency 003 and are extremely out of date when you compare it with chat GPT which is GPT 3.5 turbo or GPT for the latest API from open AI debugging a long chain error is near impossible even with verbose is equal to true even when you enable verbose it still is very important very difficult if you need anything out of the outside the workflows in the documentation it is extremely difficult to hack with even custom agents so Lang chain works fine with the ecosystem that it is wrapping but when you want to do anything outside of that it is extremely difficult and that could be also a problem the extreme popularity of the long chain is wiping the entire AI ecosystem around the workflows the recent releases by hugging face and openai re-contextualize themselves around Lang chain it's just magical Aid to the point of hurting development and crude Clarity when you have a lot of abstractions of this it's very hard to understand what is happening and if you have to understand what is happening you have to really dive into the code base to understand what is happening that is something that not a lot of developer would want to do and that's exactly the point here like if you are a developer who can get into the code base of Langston and understand it probably you shouldn't use lamp chain but if you are not somebody who can get into the code maybe that's where you can use Lang chain because you don't get into the code part in itself arvind srinivas who is the co-founder or founder of perplexity which is just like chat GPT with internet so also mentioned that they don't use land chain so perplexity does not use Lang chain that's one thing and we have never and will likely never use lime chain and then the reason that they have highlighted about why they don't want to use and channel why they are not going to use land chain starts from the sub optimality of the framework in itself models changing quite rapidly and because models change quite rapidly everything needs to be adapted or changed quickly debugging is quite important so when you have a production level software sometimes things go wrong or most of the time things were wrong and when things go wrong you need to be able to debug them in a fast and agile manner the problem is also with customization and ultimately using long chain has not given them any clear performance or abstraction benefits when you do not use Lang chain also and that's that's one of the reasons why they are not using Lang chain but again going back to our same point if you are that developer who can actually understand things Deep dive then probably you do not need land share and then multiple voices are voiced out the same thing for example Jim fan said that hackability is the number one important feature to Cutting Edge AI research and products libraries that argument llms with Vector dbe search India Predator are more useful than wrappers around them so Lang chain is fantastic for Education uh and will established workflows but that work out of the box but you are better off building your own pipeline if you do not want to use slime chain because especially if you're going to build production level pipelines maybe land chain is not necessarily or maybe lunch and should not be something that you should rely on because it again does a lot of things on top of a lot of libraries that it cannot ultimately control and that's one of the problems that you would face like for me especially when I make YouTube videos I make on top of lot of Open Source libraries and sometimes those libraries make changes and then my viewers comment comment that this function has been changed can you update the course which is not necessarily possible for me to do it so the same problem exists on long chain as well because long chain is built on top of lot of these libraries whenever any small changer and of course this is a very fast moving industry and a lot of changes happen most reminds of tensorflow which is like again like a bulk of lot of these wrappers and libraries all together and that's one of the reasons Community never accepted tensorflow as one stop solution and then they always fell back with pytosh which provided them flexibility and lot of these Deep dive access like you can dive into the code and build anything that you want ultimately this is not an issue because the founder of long chain hydration Chase is quite open about the feedback that they are getting and also acknowledge the fact that a lot of things to have to be Rewritten and that's a good thing for the community especially when you're sharing comments with good faith so let's see if Lang Chen will get transformed or not because they've raised a huge amount of funding recently and they are also one of the poster boys of the modern llm world but whether Lang chain gets some change whether langshan doesn't get some change if you ask me question personally if you should use long chain or not the answer is quite simple if you are new to the llm world if you are new to the AI world I don't think there is any harm in using Lang chain like if you want to build a quick and crappy product or project I think Lang chain is still one of the the easiest way that you can get started with in fact like much easier would be to use Lama index but if you are trying to build or move the same code to a production level code maybe that is the time that you should start reconsidering your entire Pipeline and what kind of technical debt that you want to leave behind and maybe at that time Lang chain is probably not a good solution for you to go with because now you have to fix a lot of things in like your pipeline to work efficiently and also for your pipeline to have the ability to change things when things change under the hood overall I love Lang Chen I have been a strong advocate of long chain but this problem exists and that's why I wanted to make this video if you have faced a similar problem opinion about Lang chain please let me know in the comment section I would like to hear from you