langgraph/tutorials/rag/langgraph_agentic_rag/

[giscus[bot]]

langgraph/tutorials/rag/langgraph_agentic_rag/

Build language agents as graphs

https://langchain-ai.github.io/langgraph/tutorials/rag/langgraph_agentic_rag/

[andresucv96]

Hello,

Is it only me that whenever the decision is to reformulate the question and the state goes back to the agent, it stops there. It is not going back to the retriever to get more and better context...

[hinthornw]

By any chance, do you have a LangSmith trace you could share to illustrate the issue?

[100stacks]

Hi @hinthornw,

It appears this warning in #722 (comment):

USER_AGENT environment variable not set, consider setting it to identify your requests.

was introduced with this community: add user agent for web scraping loaders PR. This PR is referring to User Agent in a HTTP Request Header to support web scrapping use cases for LLMs.

Unfortunately, this warning is rendered whenever anyone imports and uses langchain_community.document_loaders that calls
user_agent.py. Btw v0.1 docs reference web scrapping directly, though Idk if the page was ported to v0.2 docs. 🤔

I don't think this environment variable is needed for a majority of LangChain use cases, and recommend it be made an optional environment variable or simply do not render a warning if USER_AGENT is not present.

[bnbabu55]

@andresucv96, I am facing the same issue, I see that Rewrite is setting the end condition, not able to figure out why.

[bnbabu55]

@hinthornw, here is my langsmith trace for the issue above. https://smith.langchain.com/o/8045ed60-fa03-5794-b09c-62794425fd1a/projects/p/f7928a5e-dad0-4325-9fde-8ae6719672aa?timeModel=%7B%22duration%22%3A%227d%22%7D&peek=eca9a236-582b-48bf-b8da-d55ee786fa8a&tab=0&Volume=Success&Latency=P50&Tokens=P50&Cost=P50&Streaming=P50

[raselai]

Whenever I run this Agentic RAG I am getting this error bellow:

"USER_AGENT environment variable not set, consider setting it to identify your requests."

[Mooo49]

I face the same issue wile using streamlit , but it works well on google colab . so we need a solution with streamlit ??

[bnbabu55]

I am just running this as a standalone python script, even then this is happening.

[Haopeng138]

export USER_AGENT=myagent  

This work for me

[Muyuu77]

import os
os.environ['USER_AGENT'] = 'myagent'

Maybe try to put these lines at the beginning of your python file.

[pengyulong]

Setting "USE_AGENT" like this seems to be of no use.

[Dumplingisabeast]

when defining the state class, when would you use:
Annotated[Sequence[BaseMessage],...] vs. Annotated[list[AnyMessage],...]?

[pguatibonza]

Hello! I want to build an agentic rag with other tools as well, but I dont know yet how to handle it. For example, If I have another tool(i.e web search), how can I ensure that it follows another path, instead of the rag path(grade, rewrite, generate)

[woodswift]

format_docs() is defined, but it is not used? Could you please explain why?

[sreedevi1249]

create web_search node and in the grade_doucment node if documents are not relavent then you can return web_search instead of rewrite that way you can achieve adding one more tool web_search

[Jimmy-L99]

I use my own deployed llm and db, meet error as follow:

AttributeError: 'NoneType' object has no attribute 'binary_score'

[yuehua-s]

You can try replacing score = scored_result.binary_score with score = scored_result["binary_score"], which works for me.

[ricardo-casa]

what benefits does this architecture have over adaptive rag?

[sreedevi1249]

i tried little different way where if the agent is unable to answer the question meaning out of context we can include the web search which is almost similar to adaptive rag

[Meshinfo]

For all the examples, the index had already been created and did not require updating. However, in a real chat assistant scenario, the index is only created after the user uploads files. It is only at this point that the retrieval tool can be instantiated. Do you have any idea to handle such situations?

[itsihsan]

The following two issues need to be addressed within the code:

Handling Compound Questions:

1 - When a user asks a compound question (for instance., What is an agent, and what is prompt engineering?), the Agent node splits the question into multiple (in this case two) sub-questions (e.g., What is an agent?, and What is prompt engineering?).
2 - The retrieve node retrieves documents for each sub-question, resulting in two tool messages (each containing the retrieved documents for corresponding sub-question in their content part).
3 - Currently, the grade_documents node is grading the contents of the last message only, i.e., the retrieved documents from the last tool message (ignoring the documents that were retrieved in the earlier tool message(s).)

last_message = messages[-1]
docs = last_message.content 

This content is then graded against the original compound question in the grade_documents node, i.e., question = messages[0].content.

I believe the following modifications will address this problem for compound questions:
4 - Ensure the grading logic handles multiple sub-questions. This means to generalize the code to ensure that documents retrieved as a result of the retrieve node for each sub-question are graded against their corresponding sub-question in the grade_documents node.
5 - If all retrieved documents for the compound question are deemed relevant, update the generate node to consider all documents, and not just the ones from the last tool message. Currently,

messages = state["messages"]
question = messages[0].content
last_message = messages[-1]
docs = last_message.content

Issue with Memory Checkpoint:

1 - When a memory checkpoint is added to the graph as below,

from langgraph.checkpoint.memory import MemorySaver

memory = MemorySaver()
graph = workflow.compile(checkpointer=memory)

the question, i.e., question = messages[0].content will always reference the first question in the conversation, instead of the most recent user question. This causes the rewrite node to engage incorrectly due to the outdated questions.

This can be addressed by capturing the user's question and storing/calling it appropriately as below,

class AgentState(TypedDict):
    messages: Annotated[Sequence[BaseMessage], add_messages]
    user_question: str

def agent(state):
    print("---CALL AGENT---")
    messages = state["messages"]
    user_question = messages[-1].content
    model = ChatOpenAI(temperature=0, streaming=True, model="gpt-4-turbo")
    model = model.bind_tools(tools)
    response = model.invoke(messages)
    # We return a list, because this will get added to the existing list
    return {"messages": [response], "user_question":user_question}

Calling the stored question in the generate node as below:

def generate(state: State):
    print("---GENERATE---")
    question = state["user_question"]
    # The rest of the code stays the same

I would appreciate feedback from anyone who has addressed the first part. If so, please share the updated code.

[kishorekkota]

Trying to implement this with Azure Search and Azure Hosted Open AI. My retrieval call is not being invoked at all and straight getting responses from Open AI vs sourcing from Vector Store.