langgraph/reference/checkpoints/

[giscus[bot]]

langgraph/reference/checkpoints/

Build language agents as graphs

https://langchain-ai.github.io/langgraph/reference/checkpoints/

[WesGBrooks]

Love this doc and the concept of checkpointing -- when will support for more reliable persistence methods be added? ie. updates to langchain-postgres re: https://github.com/langchain-ai/langgraph/blob/main/examples/persistence_postgres.ipynb

[WesGBrooks]

Actively building on top of langgraph right now for a spiritual mentor app, so will let you know how it goes as I try it! This looks like exactly what we need to enable a long-standing mentor agent! Thank you for sharing.

I had just been considering how to deploy a langgraph via API with streaming this morning considering that would enable me to build a headless agent backend so we can deploy our agent anywhere (streamlit UI, discord, whatsapp, etc). So this is timely!

[elvenking]

@WesGBrooks Exactly ! I feel like a lot of examples on net showcases LangChain/LangGraph abilities in a more monolithic manner (e.g. NextJS app where the langchain is directly tied in RSC / Api Handler etc). Real world require separation of concerns and separate deployment of "complex, headless AI backends". I am currently trying to achieve the same with LangServe + LangGraph deployed on Koyeb and consume lcel stream_events in NextJS app using RemoteRunnable.

Would be really happy to see some nice examples & best practices

[mathlover777]

@elvenking How you are hosting langgraph with langserve, After a lot of struggle i was able to finally dpeloy langchain agents with langserve properly. langgraph seems much better from developer point of view. Can you share how you were able to get it ? and also does streaming work ?

[elvenking]

@mathlover777 sorry for late response. I think the best DX your could find is in LangGraph Cloud. There is a nice video explaining how to set it up: https://www.youtube.com/watch?v=EKNiz_fWrDk

[TonySimonovsky]

good q. probably sometime in the next month or two. we're currently focusing a lot of efforts in https://github.com/langchain-ai/langgraph-example - which besides a more persistent (postgres) checkpointer, offers things like background runs, configuration, and other functionality. we'll probably publicize this more heavily also in a month or so, but if you want to try it out before then we'd love feedback!

Any update on that?

[elvenking]

@mathlover777 I have a LangServe API deployed on Koyeb but currently without LangGraph - just an AgentExecutor. Then in NextJS i am using RemoteRunnable pointed to mentioned backend. With RemoteRunnable you can use the same methods as on Runnable, e.g. streamEvents()

[mathlover777]

I am able able to deploy Langserve in AWS (ECS with CDK, and very similar to Koyeb, ie each deployment will give a backend url), but looking for resources to do similar things with langgraph, as here the development experience seems nicer.

[mehdihosseinimoghadam]

Hi, thanks for your great work
What will happen in the following scenario if I use checkpoint:
Here is the scenario:

User starts with an initial prompt and then agents will work one after another (plan -> tool_execution-> tool_execution_route) and for some reason it will break at tool_execution_route,
now if I use the checkpoint will it start from where it left or the pointer will point at start.
(What I mean in general is, when we run the graph if for any reason process craches and then we load the checkpoint will we start from were we left, if yes how? cause when I run the checkpoint it points to the starting point but it has all conversation history)

[duob-ai]

I there any way to limit the number of used checkpoints in a graph (e.g. only use the last 3) when using BaseCheckpointSaver / MemorySaver?

Trying to follow the tutorials here but want to limit the used checkpoints in the memory:

Example:
memory = MemorySaver()

tools = [basic, internal]

agent = create_react_agent(llm.with_config(
{
"run_name": "Get Items LLM",
"tags": ["tool_llm"],
}), tools=tools, checkpointer=memory, state_modifier=system_prompt, interrupt_after=["tools"])

[3x10RaiseTo8]

I'm trying to use basic sync SqliteSaver by providing the connection object created by sqlite3 lib as argument to SqliteSaver class. It's throwing this error.
ProgrammingError: SQLite objects created in a thread can only be used in that same thread. The object was created in thread id 140639518284672 and this is thread id 140638865589952.
what happened here?

[erankirsh]

Assuming you are not using it in production you can run sqlite without thread checks:
conn = sqlite3.connect('database.db', check_same_thread=False)
This will resolve the issue.
For production you should use some sort of connection pooling. Either for sqlite or better yet, Postgres or Mongo. This is well documented here:
https://langchain-ai.github.io/langgraph/how-tos/persistence_postgres/