NEW LangChain Expression Language!!

Colab: drp.li/e9trR
Blog: blog.langchain.dev/langchain-...
My Links:
Twitter - / sam_witteveen
Linkedin - / samwitteveen
Github:
github.com/samwit/langchain-t... (updated)
github.com/samwit/llm-tutorials
Timestamps
00:00 Intro
01:03 LangChain Expression Language Blog
01:40 LangChain App Learn the Language through Chat
03:15 Code Time
05:34 Bindings
06:31 Adding OpenAI Functions
07:47 Function Output Parsers
08:39 Retriever
13:05 Tools
14:00 Arbitrary Functions

[@georgesanchez8051]

You, Greg Kamradt, and James Briggs (and echohive, whom I think is extremely underrated) are such invaluable resources for this community. The best part about getting new ground-breaking features (besides having new things to try) is the anticipation for your next video where you break it all down

[@guanjwcn]

I totally agree. But I didn’t see Greg publish any in recent weeks.

[@micbab-vg2mu]

We follow the same KZbinrs :)

[@andrewlaery]

100% agree (Sam, James add so much value in this space) Greg is a new one for me… Thx! Also really enjoy Part Time Larry (some great use cases) and Patrick Loeber (good at explaining basics), and Sentdex (generally)

[@ChrisadaSookdhis]

This is quite a change! I particularly like how retrieval can be written now with the passthrough. I’ll have to rework my code, but seems worth it for the ability to do more complex stuffs and keep code understandable. Thank you for another great video!

[@marcova10]

it is scary to upgrade the library and see all the code crashing, the amount of reworks can be massive

[@christopheprotat]

You make my day. This update comes at the perfect time. Thanks for the cookbook. You create my homework for this evening. I will test it on hf models

[@alexanderroodt5052]

Using similar techniques I am converting raw scraped data in to usable form at about 5000-10000 words per hour. Its so awesome!

[@micbab-vg2mu]

The update looks promising - thank you for the video.

[@guanjwcn]

Maybe I got too much used to the old way of langchain. I feel it’s quite some big change in this update. For example, putting prompt as an input parameter is easier to understand as opposed to the pipe command in my view.

[@parasetamol6261]

Thank you. for everthing.

[@jeffersonhope2668]

Many thanks for this excellent update.

[@AdrienSales]

Piping is so elegant !

[@uniqued4ve]

very good examples.

[@at_sofdog]

Legend. Thank youu

[@zaursamedov8906]

Again great tutorial. Can u do it one on aws falcon lite please?

[@Diego_UG]

I am interested in the part to see how now it will be asked to respond to a json, and that it delivers the expected result, I am using create_tagging_chain_pydantic, so what would its new use be like?

[@paulocoronado2376]

Awesome, Sam! This is very interesting, however I didn't find a way of setting an agent with multiple tools (usually with a description) and let the agent decide which one to use. 🤔

[@samwitteveenai]

I think they are still thinking how to handle Agents with the LCEL, currently its not supported, but I expect it will be soon

[@robcz3926]

pls can you make a video about deploying an open source LLM as an API service.not necessarily building one but at least a walk-through of what one has to consider in terms of computing power for serving the model and handling multiple concurrent requests etc...much appreciated

[@simasjanusas1766]

You could potentially use a lambda function (cloud function - gcp), which would combine request input with the LLM stored on some cloud storage, to run inference against the input. Storage is cheap, compute would be only on demand, and cost lower. Unless I am missing something?

[@clray123]

@@simasjanusas1766 Yes, you are missing the model loading time (need to keep model warm in GPU memory) and the need to batch together inputs if you want acceptable performance and hardware cost at scale. See vLLM, Ray.

[@clray123]

This whole langchain fiasco reminds me of Object Relaltional Mapping (ORM) layers. Abstracting away the thing which is not really helpful to be abstracted, caused all sorts of problems in the process (you HAVE to look under the hood to understand what is going on, so hiding stuff under fancy shmancy APIs and auto-generating pieces that you would rather avoid auto-generated are not helpful). Piling up higher is not going to solve this, the solution is to get rid of the useless proxies and KISS.

[@Ascended23]

So far I agree with you (as someone who never liked having SQL queries abstracted away from me.) I use libraries when it's doing a thing that would be extremely difficult or impossible to figure out how to manage on my own. That's why I'll use something like Vue or React, so far I'm not seeing the need here.

[@HostileRespite]

We should form a group, get tatoos and call ourselves the Lang Chain Gang.

[@JonRyan-yv3oc]

would it be possible do you think to make a custom agent like react using lcel? would love to see that

[@yayingzhang1506]

Thanks for the explanation! I get that they are trying to make it simpler, but I still feel it's confusing. Still don't feel I can program freely any way I want....

[@haristan1960]

nice walkthrough!. can you compare it with Microsoft guidance ?

[@samwitteveenai]

guidance is quite different. I should make a video about though thanks for reminding me

[@____2080_____]

Although I am extremely happy with the evolution of the package and program, me with my product management background and innovation cap on, Saul Lang Chang to be very interesting in a sense that it totally disrupts the paradigm of programming. I’m beginning to suspect that the criticisms that I’ve seen across the way literally comes from developers, fearing that disruption and not being able to adapt, and to take LangChain’s lead than any real concern over not understanding it.

[@zalorthethird]

What I don't understand is at around 11:11, how does the retriever know what question to query the vector database for? I don't understand the, {"context": retriever, "question": RunnablePassthrough()} syntax. I kind of get the RunnablePassthrough. It's just what you are passing in when you invoke the chain and that is passed along to each step in the chain, right? I don't understand how the retriever knows what the question is.

[@olb47]

Is this the newer way of making chains and 'old' way of chaining will be deprecated?

[@samwitteveenai]

not yet but possibly in the future, I think they will wait to see how popular it is

[@qingtian1691]

I did this change, and it did not work, could you explain why? # chain = { # "context": itemgetter("question") | retriever, # "question": itemgetter("question"), # "language": itemgetter("language") # } | prompt | model | StrOutputParser() chain = { "context": retriever, "question": itemgetter("question"), "language": itemgetter("language") } | prompt | model | StrOutputParser()

[@piotrecode]

Do you know how to send soruce of file with that? i also need 'refine' option here. Cannot find it in docs chain = { "context": itemgetter("question") | retriever, "question": itemgetter("question"), "language": itemgetter("language") } | prompt | llm

[@samwitteveenai]

What do you mean by "send the source of the file" ?

[@piotrecode]

Please give me link to this blog post :)

[@samwitteveenai]

Hve added it the description as well thanks for reminding me. blog.langchain.dev/langchain-expression-language/

[@regularSenseAppeal]

I don't understand 6:23 at all. Your prompt at this point is "tell me 3 interesting facts about {subject}" then you go straight to : functions_chain = prompt | model.bind(function_call={"name": "joke"}, functions=functions) I can't see how that prompt about 'tell me 3 interesting facts...' is in any way related to the joke function? Could anybody elaborate please?

[@stanTrX]

My brain burned

[@heythere6390]

it's still convoluted as hell. like they just want to throw names and terminology at you all the time.

[@hansofmadata3565]

It’s been a frustrating year trying to build production apps in this space, but that’s what we get for trying to be early 🤷‍♂️

[@clray123]

They are making a complicated DSL for a simple task (parsing and passing around pieces of text).

[@NoidoDev]

@@clray123 Pretty sure this is being done with the future in mind, the idea is to build on it.

[@knoopx]

meh, not sure i like it, i guess i will have to try first but personally i think a proper templating language would make much more sense.

[@samwitteveenai]

got me curious. I have been working on a project like this, what would you like to see in a "proper" templating language?

[@knoopx]

data binding and flow control mostly, essentially a textual way to define a full chain without having to manually glue the inputs/outputs of each individual step through python. microsoft/guidance follows that approach. for a pure python DSL i would prefer something along the lines of prefecthq/marvin.

[@Jonathan-rm6kt]

Hey Sam I appreciate your videos, but I have to say that Langchain throwing in this new syntaxt has not helped at all for me. Most of the documentation and guides has not been updated to show how it's used (anything beyond toy examples), and I've found that the syntax provides for even *less* introspection. For example, when askign the docs how to inspect {context} from the retriever, it suggests to create a function and pipe it in as a RunnableLambda. I actualy like the old syntax better, because at least I could logically inspect the inputs and outputs. Now everything is hidden even more! Wish me luck...

[@samwitteveenai]

I have been playing with this a lot more myself, so I probably will make some more explainers. Hopefully that will be useful.

[@Jonathan-rm6kt]

@@samwitteveenai That would be amazing. If you want, I can share the research/questions I've done so far, maybe that can help you structure the video to answer questions for beginner/ intermediate level people like me. Thanks again for publishing all your great work.

[@AngusLou]

I just think Langchain is getting more complicated.

[@alx8439]

I think the value of LangChain is overrated. To me it's nothing more than just a wrapper around your own prompts to any LLM. There's nothing in there you cannot do yourself with few lines of python code

[@jimjones26]

Go ahead and show us how it's done!

[@knoopx]

@@jimjones26 from transformers import pipeline generator = pipeline('text-generation', model='gpt2') results = generator("langchain is", max_length=10) print(results[0]["generated_text"])

[@NoidoDev]

It's a step towards agents which seek the knowledge they'll need and responses on their own.

[@adityagaurav2816]

I think they Made it more shitty with this update ...

[@samwitteveenai]

curious why? you don't like the declarative way? I don't think it is perfect, but I do think it is a step forward in making it easier for new people etc.