I actually wasted a couple of months using Langchain until I realized all of this myself, the dev experience is just terrible and why? all we are really doing is putting some strings together and sending them to an API, there is no reason for all these stupid abstractions and having to conform to these objects. The main reason to use Langchain is so that you can swap LLM´s easily, but at the end of the day it would be much more time efficient to just refactor your code to adapt to the new LLM. Makes no sense. You know what I do like about Langchain? there will be hundreds of PM and Architects that will see the benefits of langchain on paper, and will decide to adopt the tech, without ever having tried to make something in it, it will lead to smaller teams being able to deliver more quickly and compete better.
I cannot think of a single usecase for API users with this pricing... It needs to get about 4x cheaper to enable some usecases and about 100x cheaper to make sense for many projects.
I'm at the initial demo portion rn (paused at 2:14) and im noticing a lot of cuts after you finish talking which feels like it ruins the point of a demo of a realtime ai
Hey Lewis. Nice second channel. I’ve been watching you for the last two years (ever since Ive started my comp science journey - 1 more year to go for graduation 🙂) Great video!
I wish I would find that video earlier, it would have saved me at least a couple of days and then something that was a problem for me is about the fragility of the offered solutions, a gazilion ways of doing something just so that when you need to add something more to your project it will fail, which then leads to you wasting time on the things you just learned. If you are trying to develop incrementally, as usual, it can get really bad
"Lanchain is trying to teach you langchain..." I think for me this is the biggest let down of it all. when I use Langchain I always feel as if programming is regressing to old ways which is a shame because LLMs are supposed to be removing the budden of coding. I feel sad and awful for commenting badly about Lanchain becuase I think they had good intesions. But I keep thinking that they make things look way harder and complicated than they need to be with all the abstractions. And I agree with your sentiments towards the docs, having worked with native SDKs for OpenAI, Anthropic etc. It was never this 'complicated;, I think their problem was trying to do too much in one library, and now they cant stop due to compatibility considerations, its sad
Man, your video really resonates with me. Ten layers of useless abstraction around a prompt augmenation like "...answer in json fomat, according to the following schema...".
Nice video :) I heard from many sources that LangChain is not the best fit especially for production or even learning. Have you already checked out the Agent-Swarm framework by VRSEN?
I had the exact same points and was very concerned as I felt that this way of coding is very unnecessary but didn't have the guts to speak up thinking am I the only one who has such thoughts or may this thing is something way advance that I am not able to get so speaking up might make me sound stupid. Thanks for bringing this up.
You ever use anything other than simple prompts? Agents, rag, function calling, tool calling? If not, then you don't need langchain. Don't bash something because you didn't take the time to understand it.
I agree with certain things, and disagree with most. Documentation for langchain is aweful. LCEL and chains make no sense. Memory management is a headache. But what i love about langchain is thats its very easy to build and prototype agents and multi agentic workflows. It took me too much time to learn LCEL but once i got the hang of it, along with langgraph, langsmith, langchain is unbeatable
I absolutely agree, the comparisons that are given at the end of the video and the purpose of the langchain is completely different. There are tons of helper tool startups coming up to the world of llm. Many of them you can try and integrate in langchain very quickly. Also agreeing with the difficulties mentioned in the video however clearly the point of the langchain ecosystem is quite different, I think Lewis is missing that.
@@fantinigerman very little low level control. Good for existing apps and features but to define anything important, implement custom logic, functions, anything, very little customisability. For control freaks like me, I like implementing everything from scratch and langchain is a good platform
Langchain is open source, you can always check the source code. No one is stopping you from doing that, to understand how it is working under the hood. Abstractions are good, if you know how they are working. So I encourage you to do so.
Of course! That's why I said it's no disrespect whatsoever. The abstractions mixed with the poor documentation is what made me not use it anymore. Other libraries make it easier.
Dude, Langchain is not only about calling different LLM apis... It is about chains, hence the name tell is tall Lang-chain. You can place any logic in-between.
I also hate that the api documentation is seperate from the documentation you showed. It took me months to find out it even existed because it is barely linked to from the 'main' docs. Gives me a headache
I enjoy more this kind of content which has more detail than a "understandable for all" content. Don't get mi wrong, your main content is super good too, but as a Engineer I found this more interesting.
I am waiting for ai to code ai. Do I care what abstractions that code would use. I really don’t and shouldn’t- cuz bugs (Performance or whatever kind of bug) also will be fixed by ai🥹
I have rarely returned for for a second comment but I have an interesting story to share.. One of my PMs was worried about our project reaching Xlines of code already? Even after using langchain.. Turns out it was because a team had USED langchain, that the codebase exploded.. and no cap.. she went on to replace so much stuff with ternaries, f string, and simple try, except finally.. and legit the code base was like 60% of what it was, and most commits looked like -xyz removed 45 lines of code- "optimized suboptimal code".. and I was legit lmao.. 😂😂
Same.. tried to use it 5 times, broke it 5 times in production. It's like what I call 'junior dev baity" vibes.. Edit: this content is so much fun... I'll help you shit on some more dumb techs lu.. Go for vector databases next... They can be replaced by sqllite and simple hashing.legot no cap... anecdotal but true agein.. calling yourself production ready don't make one.. production ready .. Then you can go on for the new "agentic" frameworks.. Stitching two llms half-heartedly and running semicooked daemon threads don't make you "intelligent".. (hello autogpt) .. or faking your vids (hello Devin).. I had to literally invent a language just to get a simple internal audit tool working with 3 instances of gpt4 working and then too half of the time it's blah!(According to auditors).. so no job replacements there .. at least.. one would think, auditors would be the first to get replaced because all they do is HUMAN RAG... And that came from top ISO 9001 auditor in Norway..
i took a deep dive back in October 23, then gave up because of the amazing documentation and can't get the "langchain" way of doing things (like not using f-string). 3-4 months later, codefluencers promote it heavily, thus i installed the latest version and most my code didn't work because now we have new syntax with LCEL (langchian expression language). Imagine you used LangChain in production! LCEL is the real KO for me and i never come back.
