Build a LangChain App [Tutorial]

[@BillAugersdca]

One of the most under-rated channels on YT. Thanks for all the great info Richard!

[@lucidateAI]

Thanks Bill. Greatly appreciated. Seemingly not under-rated by you!! With thanks!!

[@lordmelbury7174]

@@lucidateAI I'll second that, great video, much appreciated.

[@lucidateAI]

Greatly appreciated. Terrifically "good for the wheat" as well! Still got the bricks in the briefcase? (One of my favourite FT episodes!). Lucidate.

[@AndrewLiu1]

agreed! the teaching style is amazing!

[@lucidateAI]

Thanks Andrew. Greatly appreciated. Lucidate.

[@Indrikmyneur]

Well done. I like this format: real problem > well chosen tool > real code > explanation. Thanks.

[@lucidateAI]

Thanks HFW. Greatly appreciate the comment and you sticking with the channel after some early disappointment. I’d probably say “early prototype code”. It clearly lacks any robustness or notion of extensibility, but I hope it adequately demonstrates the concept. With sincere thanks, Lucidate.

[@petrmeissner8292]

@@lucidateAI I don't actually see the prototypiness as disadvantage at all! It's a balance between the robustness and clarity. In such intros I value clarity over robustness, because I see the biggest value in an introduction of unfamiliar tools, compositions or patterns, not making a production ready code.

[@lucidateAI]

Thanks Petr, and FWIIW I agree. As an illustration of the proof of concept this is fine. My point is that for an actual application much more attention must be paid to sound principles of design. But that would make this a coding tutorial, and while I feel qualified to provide AI tutorials, I’m simply not qualified to provide similar content on coding. So a long-winded way of saying “Thanks! And I agree” !!!! Lucidate

[@ricand5498]

The best langchain tutorial I’ve seen yet. The graphic on the parts of the library helped understand the structure of tools/capabilities which in a way I’ve never seen before!

[@lucidateAI]

Thanks. Glad you found it useful.

[@mikecy2899]

This channel is gold and so underrated.... Thanks Richard for what you're doing, mega job, brilliantly explained (even to me, a person with no developer experience)

[@lucidateAI]

My pleasure! Glad you found it so useful. Lucidate..

[@BlakeAlexanderGames]

I’m always ready for the next video

[@lucidateAI]

You certainly are!

[@yxs8495]

You beat me, I wanted to be first.

[@lucidateAI]

Next one should be out Sunday afternoon UK time. You heard it here first!

[@lucidateAI]

Did you enjoy the other videos in the LangChain series? kzbin.info/aero/PLaJCKi8Nk1hwZcuLliEwPlz4kngXMDdGI

[@LaFragas]

EPIC EPIC EPIC EPIC!!! this is the content I was looking for. for dummies epic for starting in the space.never felt more compelled to start doing something, thanks! LFG

[@lucidateAI]

Really glad you found it so useful and inspirational! Have you had a chance to check out some of the other videos in this playlist-> kzbin.info/aero/PLaJCKi8Nk1hwZcuLliEwPlz4kngXMDdGI. Or in this where we look at what others have build using LangChain components? -> kzbin.info/aero/PLaJCKi8Nk1hy48P5fJWkD-AJBVi9JDhTZ

[@LaFragas]

@@lucidateAI just found your channel and will have a look, thanks for sharing!

[@lucidateAI]

Great. Looking forward to hearing your feedback on the other videos. Let me know what is missing and you’d like to see.

[@lucidateAI]

We’re the playlists useful? Lucidate.

[@WeirdoRocks]

Richard, thanks for this great tutorial. Loving it and will play around with your code. :)

[@lucidateAI]

Thanks Weirdo! Greatly appreciate the positive comment, (appreciate all comments, but have a proclivity for the positive ones!) I hope you find the code insightful. Lucidate.

[@gui_p_mello]

Really good video! I hope you make many more

[@lucidateAI]

Thanks Guilherme for your comment! Greatly appreciated, I'm glad you enjoyed the video. You'll be glad to know that there are many more in the pipeline. Please check out the rest of the Lucidate channel -> www.youtube.com/@lucidateAI/featured. You'll see a series of popular videos and videos organised into playlists on various topics. I'm keen to make the subjects as relevant as possible to topics that currently interest viewers, so please let me know what you feel I could cover, or if there are subjects you feel I could address in a different way. Appreciate your comment and support of the channel. Lucidate.

[@lucidateAI]

Have you had a chance to look at any of the other videos in the LangChain series kzbin.info/aero/PLaJCKi8Nk1hwZcuLliEwPlz4kngXMDdGI Lucidate.

[@roscoevanderboom8449]

I really appreciate the videos and learn a lot. One thing I want to ask for is a Langchain video that doesn't use the OpenAI embedding. I've seem at least 10 videos, and everyone uses OpenAi. This is great and all, but I'd love to see a video where some uses a different model such as Llama, Alpaca, Vicuna. I'm not Python savvy and really have to follow videos closely to learn this stuff and I want to develop with something other than OpenAI. If the process is the same, let me know.

[@lucidateAI]

It is a good point and I'll try to address it in future videos. Part of the reason that OpenAIs GPT-X.Y models are so used is that they are: good (frankly very good), stable, have well-documented APIs (ways to access them without a browser as in the last part of this video) and pervasive - everyone has access to them. But you are 100% correct, as content creators we need to do a better job of showing how there are other tools available. Indeed, one of the major selling points of technology frameworks like LangChain is the ability to interface to LLMs plural. There is an 's' after 'LLM'. Probably in my case, though I'm ashamed to admit it, there might be a bit of laziness too. OpenAIs GPT series 'just work' so why make it harder? We'll youve given me a reason to make it harder - to make it more relatable. So thanks for giving me a kick to use something different. It won't be for a few videos I'm afraid as the next few are either finished - or significantly progressed. But your suggestion is an excellent one and I will act on it. But equally importantly, I sincerely appreciate your kind words and support of the channel. With thanks, Lucidate.

[@roscoevanderboom8449]

@@lucidateAI Well, youv'e earned yourself a new subscriber and I'll be keeping an eye out for new vids. I've also build my own gpt app using Tauri + React + Firebase. It works pretty well, but one thing that's bothering me is that the context limit of the conversation completion API is 4k token. This run out pretty quickly when using gpt for code snippets, meaning I have to start a new chat thread often. Currently looking into solutions like privateGPT and auto-gpt.

[@lucidateAI]

Thanks Roscoe, I hope the future videos live up to your expectations. Please feel free to browse around the rest of the Lucidate channel to see if there is anything that is helpful -> www.youtube.com/@lucidateAI/featured. I appreciate that you said in the first comment that you weren't "Python savvy" so LangChain might not be the answer here, but the Indexes components in LC might help with the context limit issue you highlight above. Might you consider using LC/Indexes as an approach for this?

[@khushpatel6296]

Hello Lucidate, I am asking here as this is your latest video, I have seen your videos regarding Langchain, I am curious how to optimize inference while using LLMs in Langchain using TensorRT or Onnx Runtime... As in Industries, It's very obvious, to save the time as well as computation cost... In TensorRT while using open source models, we have techniques like quantization and a few more more for that... So In Langchain is there any way to do this?

[@lucidateAI]

Think of LangChain as a set of components that aids, amongst other things, prompt engineering and integration of single or multiple LLMs with other applications and information services. This is clearly not the only way to think of LangChain, but it works for me. Once you have built and tasted your chain, why could you not deploy it to TensirRT/Onyx by following these instructions-> docs.nvidia.com/deeplearning/tensorrt/quick-start-guide/index.html. This is not something I’ve tried, so it is something I’d really like to know (ie, this is not a rhetorical question).

[@Dr_Tripper]

I have been at this for about 5 weeks now and I am finally seeing the light! Now I am understanding the code (haven't coded in 30 years) and soon I will have a working app. I have come to the conclusion thought that OpenAI dominates most everything. SO, being a rebel, I have only been focused on how to have my own AI without paying - PRO TIP: if you don't want to lose sleep, hair, patience and weight, pay someone else! :) But, but now I am right on the verge of having an offline (online when given permission) personal AI. When I get my GPU that is. :( What a time to be alive. Lucidate, Witteveen, Berman and (edit)Renotte

[@lucidateAI]

Thanks Rick. Best of luck regaining sleep and hair! Glad you are finally seeing the light. Keen to hear how things turn out once you have a working app. Experience suggests that this is when the fun really begins! There are always things that don’t become apparent until you fire things up in anger for the first time! But don’t be deterred!! Do you mind if I ask what your specific use-case is? Agree with your assessment of the other KZbin channels. I’m humbled and honoured to be mentioned in the same sentence. Hopefully our offerings complement one another and where one of us is a little weaker the gaps are filled in elsewhere. Many thanks for your comment, and as I said - keep us all updated with your progress. Lucidate.

[@Dr_Tripper]

@@lucidateAI Glad to be in spectator of the bleeding edge. My use-case is of course a Jarvis like assistant. Perhaps if someone want me to build something for them I will but I just enjoy the flow. See on the next one!

[@lucidateAI]

Thanks Rick. Appreciate you sharing the use-case. I assume you are waist deep in LangChain Agents?

[@Dr_Tripper]

@@lucidateAI I am at the gate. I stopped short to learn about pipelines and transformers - and I am working my way through them to see what each do and how I can use. I am waiting on a GPU accelerator (I can only afford a K40 :( ) so I can have CUDA (As you know CUDA is essential for some tasks as I cannot get ROCm to work in WSL although Stable Diff uses my AMD GPU curiously). All the while I am looking for a pre-trained (yet trainable) model that I can start building my base on. Any suggestions for a multimodal enc/dec model with ~10,000 tokens?

[@lucidateAI]

Hi Rick. Best of luck in your endeavours. The first observation is 10,000 tokens is an extremely small dataset and there is a huge risk of any model even with a modest number of parameters overfitting. You can take a look at some of the 'tiny' models on HuggingFace as a starting point huggingface.co/models?sort=downloads&search=tiny but these are still very large relative to your training set. Have you had a chance to look at any videos in this playlist? -> kzbin.info/aero/PLaJCKi8Nk1hzyMP062yzf9GMWiUYihZ4g There are some pointers here that might help you. And there may be some useful content here too -> kzbin.info/aero/PLaJCKi8Nk1hzTXRS_w7IeH6UZdtYvpHP8. Might you be able to get a larger dataset for training? This is the definitive paper on giving you a starting point for model size relative to amount of training data -> arxiv.org/pdf/2203.15556.pdf Don't take it as gospel truth, but it will provide you with a great place to start sizing your model. I hope this helps. Please feel free to ping me with more questions and I'd love to hear updates on your progress. If in the future you wanted to collaborate on a video journaling your progress I'd be open to that. Best and thanks Lucidate.

[@joser100]

Hi there, can you fix the github link? (for others... ) the end-parenthesis came inside the link, which makes the direct access fail... Other than that, congratulations for another great video... from a solid follower of your work, love each video... looking forward to the next one.

[@lucidateAI]

Good catch Jose. My apologies for the snafu in the first place. Greatly appreciate you pointing this out as well as your positive comments on the Lucidate content generally! With sincere thanks, Lucidate.

[@hadijannat4821]

New gpt with 32 and 16k has been announced!

[@lucidateAI]

Thanks Hadi. Was this video useful for the project you mentioned in the chat on the Lucidate website? Richard.

[@hadijannat4821]

@@lucidateAI I appreciate your input, Richard. It was truly beneficial! I've picked up some tips and I'm planning to utilize them for my objectives.

[@lucidateAI]

@@hadijannat4821 That’s great. Keen to hear how you get on with your project. And now (or soon) you have a larger context window! Richard

[@whemmakatatt5311]

top tier

[@lucidateAI]

Thanks! Glad you found it insightful. I hope that there is other content on the Lucidate channel that you enjoy as much! Lucidate.

[@juliena6575]

Merci !

[@juliena6575]

You deserve a lot more subscribers!

[@lucidateAI]

Many thanks for your kind words and generous contribution. I hope that you find the other material on the channel just as insightful. Richard

[@lucidateAI]

That is very kind of you to say! Tell ten friends/colleagues and then get them to do the same (and... and.... and...)! My limited knowledge of mathematics says that if that happens there will be a few more subs and views ;-) Greatly appreciated!