Checkpointer (memory concept in controlled agent)

[akash97715]

Checked other resources

• I added a very descriptive title to this question.
• I searched the LangChain documentation with the integrated search.
• I used the GitHub search to find a similar question and didn't find it.

Commit to Help

• I commit to help with one of those options 👆

Example Code

My question is the checkpointer concept available for the controlled flow agents where we have a set of defined nodes and edges. Impoertantly we dont do bind llm calls(basically assistant thing)., there will be just a tools and set of defined nodes and edges. how we can have the checkpointer present there. 

because what i have observed, if we pass thread id also it doesnt work it just hit the set of defined steps and then finnish. 

please let me know if any ellaboration is required?

Description

My question is the checkpointer concept available for the controlled flow agents where we have a set of defined nodes and edges. Impoertantly we dont do bind llm calls(basically assistant thing)., there will be just a tools and set of defined nodes and edges. how we can have the checkpointer present there.

because what i have observed, if we pass thread id also it doesnt work it just hit the set of defined steps and then finnish.

please let me know if any ellaboration is required?

System Info

langgraph latest

[dosubot[bot]]

I wasn't able to find specific information on using the checkpointer concept with controlled flow agents that have defined nodes and edges without LLM calls in LangChain. The available documentation primarily covers using checkpointers with memory in scenarios involving LLMs and conversational agents ^[1][2].

If the checkpointer isn't working as expected in your setup, it might be due to the way the agent is configured or how the checkpointer is integrated with the defined nodes and edges. You may need to ensure that the checkpointer is correctly set up to handle state across the specific flow of your application. If you have further details or specific code snippets, it might help in diagnosing the issue.

To continue talking to Dosu, mention @dosu.

[akash97715]

@dosu , checkpointer is totally working fine for llmbind agnets with tools and all, but it doesnt work with controlled agent. configuration everything is perfect there is no issue, when i try to pass thread id in llmbind it works, but for controlled it keeps on running the set of defined graph nodes and edges.

[dosubot[bot]]

The issue you're experiencing with checkpointers not working with controlled agents in LangChain, despite working with llmbind agents, might be due to the inherent differences in how these agents handle state and flow. Controlled agents, which follow a predefined set of nodes and edges, may not inherently support the dynamic state management that checkpointers provide, as they are designed to execute a fixed sequence of actions.

In contrast, llmbind agents are more flexible and can leverage checkpointers to maintain state across interactions, especially when using thread IDs to track conversation history. This flexibility might not be present in controlled agents, which could explain why the checkpointer doesn't affect their execution as expected.

To address this, you might need to explore alternative ways to integrate state management within the controlled agent's flow or consider modifying the agent's architecture to better accommodate checkpointers.

To continue talking to Dosu, mention @dosu.

[eyurtsev]

@akash97715 i don't understand the question. Are you using langgraph checkpointers? What is a controlled agent?

[akash97715]

Hello @eyurtsev thanks for your quick response. Controlled agent is something where we have set of defined edges lets say self rag agent(https://github.com/langchain-ai/langgraph/blob/main/examples/rag/langgraph_self_rag.ipynb), how we can acheive langgraph checkpointers here.

we are using Langgraph checkpointer.
My usecase:
if i have used a query lets say who is Mr bennet and there is a set of steps defined to retreive grade, re-write generate etc.

my follow up question :

tell me something more about him? or what i had asked you?
then it again goes for set of steps that i have defined. is there a way it uses the checkpointer here in this controlled graph .

let me know still i am not clear. i will try to explain more.