langgraph/concepts/application_structure/

[giscus[bot]]

langgraph/concepts/application_structure/

Build language agents as graphs

https://langchain-ai.github.io/langgraph/concepts/application_structure/

[120215727]

great!

[aidanj18]

Is there a way to exclude certain files/directories from being included in the docker image that gets built. For example an excludes field as well as a depdencies field in the LangGraph configuration file?

[Juanchote]

have you tried to add .dockerignore file to the root folder?

[grega913]

Hi, I am on Windows 10. Did go through all the necessary steps mentioned in langchain-academy in this video: https://www.youtube.com/watch?v=o9CT5ohRHzY&t=149s&ab_channel=LangChain
I am in folder "langchain-academy/module-1/studio" and did pip install "langgraph-cli["inmen" g==0.1.55".
When I run "langraph dev" the app starts and I am redirected to the "https://smith.langchain.com/studio/thread/...blablabla...baseUrl=http%3A%2F%2F127.0.0.1%3A2024" and see the graph and pretty happy here. But when I submit something through input field - I get this error:
Traceback (most recent call last):
File "c:\Users\gs\Documents\AI24\LangChain_202409\langchain-academy.venv\Lib\site-packages\uvicorn\protocols\http\httptools_impl.py", line 409, in run_asgi
result = await app( # type: ignore[func-returns-value]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "c:\Users\gs\Documents\AI24\LangChain_202409\langchain-academy.venv\Lib\site-packages\uvicorn\middleware\proxy_headers.py", line 60, in call
return await self.app(scope, receive, send)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "c:\Users\gs\Documents\AI24\LangChain_202409\langchain-academy.venv\Lib\site-packages\starlette\applications.py", line 112, in call
await self.middleware_stack(scope, receive, send)
File "c:\Users\gs\Documents\AI24\LangChain_202409\langchain-academy.venv\Lib\site-packages\starlette\middleware\errors.py", line 187, in call
raise exc
File "c:\Users\gs\Documents\AI24\LangChain_202409\langchain-academy.venv\Lib\site-packages\starlette\middleware\errors.py", line 165, in call
await self.app(scope, receive, _send)
File "c:\Users\gs\Documents\AI24\LangChain_202409\langchain-academy.venv\Lib\site-packages\starlette\middleware\cors.py", line 93, in call
await self.simple_response(scope, receive, send, request_headers=headers)
File "c:\Users\gs\Documents\AI24\LangChain_202409\langchain-academy.venv\Lib\site-packages\starlette\middleware\cors.py", line 144, in simple_response
await self.app(scope, receive, send)
File "c:\Users\gs\Documents\AI24\LangChain_202409\langchain-academy.venv\Lib\site-packages\starlette\middleware\base.py", line 176, in call
with recv_stream, send_stream, collapse_excgroups():
File "C:\Python312\Lib\contextlib.py", line 158, in exit
self.gen.throw(value)
File "c:\Users\gs\Documents\AI24\LangChain_202409\langchain-academy.venv\Lib\site-packages\starlette_utils.py", line 82, in collapse_excgroups
raise exc
File "c:\Users\gs\Documents\AI24\LangChain_202409\langchain-academy.venv\Lib\site-packages\langgraph_api\base\shared\sse.py", line 36, in call
task_group.start_soon(wrap, self.listen_for_exit_signal)
^^^^^^^^^^^^^^^^^^^^^^^^^^^
AttributeError: 'EventSourceResponse' object has no attribute 'listen_for_exit_signal'. Did you mean: '_listen_for_exit_signal'?

Do you have any idea what is the problem?

[grega913]

OK, I solved this by modifying 2 lines in .venv\Lib\site-packages\langgraph_api\base\shared\sse.py:
line 36: task_group.start_soon(wrap, self.listen_for_exit_signal) -> task_group.start_soon(wrap, self._listen_for_exit_signal)
line 41: await wrap(partial(self.listen_for_disconnect, receive)) -> await wrap(partial(self._listen_for_disconnect, receive))

this solved it for me

[martinobettucci]

I am sorry but this nukes completely the usefulness this way for researchers:
first of all, how am I supposed to know what's the content of the json? why do I have to create a nth-way of declaring dependencies when I already have my environment? why do I have to specify the dependencies at all again? why my graphs are required to be a python script? Could it not run like an IPython instance inside my already running environment? I can't test run my jupyter notebooks this way, I can't use conda environment.. anyway, this is completely useless the way it is for anyone wanting to do scientific LLM as plain python script is simply not the reality.

[hinthornw]

The problemS are (there is at least one problem per line required), I DON'T HAVE a .env file because my vars are declared inside my jupyter environment or inside the secret enclave of the python interpreter which you know, I DON'T WANT to export.

1. If you are just running locally with langgraph dev, you can omit the env line. It will use your interpreter environment. The env config is used if you are running in docker.

I DON'T HAVE the graph in a dedicate python script because my agents are spread in multiple cells all around the notebook and there is no easy way to compile a stand alone python script without copy/paste back and forth between my notebook and the file which is error prone completely nuking the usefulness of debugging.

2. You could use nbconvert or jupytext to export to a python file, though you would want to remove extraneous calls. This is a server environment so it's not that common to run from a notebook. Are there other production web server frameworks that you've found work well in notebooks that you'd suggest we study?

I DON'T have a dependency list because the IPythons environments depends by the python server which are pre-built most of the time.

3. (same as in (1) Deps in langgraph dev in-mem server just use your local environment. Deps are used for docker and production deployment.

[Juanchote]

I see, let me explain to you what my workflow looks like, I see we mostly use the same tools but in a very different way. My background is webdevelopment, mostly backend/devops so I guess I am more familiar with some of the toolings related to the langgraph platform itself. I am not saying here you need to change your way of working in your development environment, just sharing how I do it which seems to be more aligned with langgraph team than your way.

I don't think having env vars hardcoded in your JN are a good practice, specially if at some point you plan to share the JN with someone, I use python-dotenv and I have in the first cell of my JN this:

from dotenv import load_dotenv
load_dotenv(override=True)

now I share the same .env between my webserver and JN without exposing the variables in case I want to push my project to github.

I think the main issue here is the mix of JN and a webserver. a webserver doesn't work with JN, JN are for experimenting fast but you cannot deploy JN to a webserver and expect an API to use the code in there. The code needs to be in python files if you want the webserver to be able to read and use the code, that is how webservers work, and it is what langgraph cli does, spins up a set of dockers with a webserver with an api for you to use.

in my case, I use jupyter notebook for assembling my workflows, classes, code, etc, so I can iterate faster the logic, once I have a workflow that works, I move the code to their py files, and they are ready for e2e testing or deployment to the cloud, but the JN are for fast local development.

What langgraph cli is doing under the hood is creating a docker container with your code in it, but in order to work, it needs to recreate your venv inside the container and the only way to do that is by a requirement.txt because is the best practice when working with docker container. It is a package manager with isolated system. It is like a VM.
At leat in conda you can do

conda list -e > requirements.txt

and it will create the requirements.txt based on the actual env, so not a big deal, just remember to run that from time to time whenever you add new dependencies.

Now, to be fair, I tried the langraph server and I don't like it because I found it too green and rigid and not good enough for the stuff I need to do and I just use my own fastapi server to run the agents.

[martinobettucci]

Thank you Juanchote, thank you hinthornw.

I get that for production you want to have an isolated environment, we also deploy in dockerized apps but during development, it is a very common and convenient way to develop your agents and llm inside notebook. The studio requires a production compliant environment, I get it, simply this is impractical because the studio is for debugging AFAIK.
Cool for the environment variables, nice to know: that should be added in the doc of course.
As per the exporting the notebook, this is also impractical, what would be great is for the studio to be integrated into the running IPython instance so we can change the code and update the references between the notebook, add breakpoints, .... while observing the graph responding in the convenient Studio UI.

What we do, for now, we implemented the listener at the langchain layer and we output everything in a file (like a logger). It works but...it is a pity a Studio Dev environment exists and do not use the dev environment.

my 2ct

[martinobettucci]

Ok, I've managed to make it work but it is a mess: the studio framework is a mess, honestly.

I've retrospected the code starting from the graph name using ASTOS and exported dynamically the code to a python script which is updated from my IPython execution engine via a listener added to the zero-mq that runs the commands between the notebook and the python instance server (this is the way jupyter actually works). In simple words, when I execute the code on the notebook and if the code is used by my graph detected by an acyclic reference of symbol extracted from the running abstract syntax tree of executed cells, it updates the python script accordingly. I generate dynamically a project folder with the init.py (another important info the doc failed to pass on how the dev server loads the graph referenced in the json. About the requirements.txt, I use the pip module api to retrieve the packages with their version to generate a file and for the .env is not required because I confirm current environment work but I have to run the langraph dev inside the notebook kernel.. but it works so its fine for this.
After that, I have wrote a flask server managing à ngrok proxy because it is IMPOSSIBLE thanks to the CORS to connect a https page like smith.langchain.com to a http pointing to local resources like 127.0.0.1 !
The flask server add the cors policy and proxy all requests back to the local langraph studio dev instance which handle the calls.

I am sure there is a better way like, letting the studio be self hosted for exemple instead of forcing to use the https://smith.langchain.com/ instance. Also a self hosted solution could "catch" the zero-mq method directly as to runs live code instead of static code from a file.

[martinobettucci]

I mean "the studio connection code I have to make is a mess" sorry, but I can't edit