Allow for routing between agents and llmchain

[thecasual]

Feature request

I would like to receive a prompt and depending on the prompt route to a dataframe agent or a llmchain. For example, the question may be about the aggregation of data in which I would like to utilize the dataframe agent. If the question is about pdf files I would like to use the llmchain to handle this.

Motivation

The ability to have a universal answer flow is ideal for my scenario. I would like the ability to handle any question and respond with either a response from the dataframe or from llmchain.

Your contribution

I am still going through the documentation and code repo itself. I can certainly work towards a PR. Want to for sure ensure that this is not already possible. Thank you!

[dev2049]

hey @thecasual! would this do the trick? https://python.langchain.com/en/latest/modules/chains/generic/router.html

[thecasual]

Hey @dev2049 , I am not sure that you are able to route to agents the same way you are able to route to chains. Perhaps I am misunderstanding this concept.

[saxenauts]

Hey, did you find something that works like this ? Do you want to work on it together ?

[thecasual]

I have not found something yet and sure would be happy to work together on this if a solution doesn't already exists.

[sk9288go]

hI! i'm going through the exact same dead end with this, the moment i found out about router chain i thought this would do the trick but haven't got so far with it... have you solved this?

[Dolmedon]

Same here, it would be super useful to route to an agent for specific data questions and an LLM for a more generic or creative conversation

[adamcolon]

This is definitely a shortcoming of the framework.

Being able to route to an agent is essential.

[wolfassi123]

I am facing the same issue. I have an agent with several routes that consist of ConversationalRelationChain and LLMChain. The first issue was that each one has a different input. I tried using LOTR (Lord of all Retrievers) to fix the issue but I have been getting the following issue over and over "Saving not supported for this chain type.". I currently am seeing no solution for it. If anyone came up with any solution for it please mention.

[aayushchhabra1999]

Need something like this as well. Did someone figure out a solution yet?

[a-a-ron-12]

I too just stumbled upon this dilemma today. I initially thought RouterChain would resolve my situation, but I then realized it would have to dynamically switch between different agents, since the chain types and configurations would have to be different between different intents.

I can share how I plan to tackle this dilemma as a workaround for the time being: I'm going to set an initial system prompt that will require a multi-classification output based on the user's inquiry. That classification output will then route the proper agent based on Python "if" conditions. It's not ideal, I know, but may help generate some ideas for others in limbo.

[rodrigocruzs]

Has anyone found a workaround? I also need this

[csprocket777]

Same, I am also looking for a solution to this but in the js/ts framework.

[hwchase17]

There are several ways to do this, here's an example using LangChain Expression Language

# Setting up SQL Agent
from langchain.utilities import SQLDatabase
from langchain.agents import create_sql_agent
from langchain.agents.agent_toolkits import SQLDatabaseToolkit
# from langchain.agents import AgentExecutor
from langchain.agents.agent_types import AgentType

db = SQLDatabase.from_uri("sqlite:///Chinook.db")
llm = OpenAI(temperature=0, verbose=True)

agent_executor = create_sql_agent(
    llm=OpenAI(temperature=0),
    toolkit=SQLDatabaseToolkit(db=db, llm=OpenAI(temperature=0)),
    verbose=True,
    agent_type=AgentType.ZERO_SHOT_REACT_DESCRIPTION,
)

# Set up the routing logic
from langchain.chat_models import ChatOpenAI
from langchain.prompts import PromptTemplate
from langchain.schema import StrOutputParser
from langchain.schema.runnable import RunnableMap

prompt = PromptTemplate.from_template("""If the question is about SQL, respond with `SQL`. Otherwise, respond `OTHER`.

Question: {question}""")

router_chain = prompt | ChatOpenAI() | StrOutputParser()

# Set up the base chain
llm_chain = PromptTemplate.from_template("""Respond to the question:

Question: {input}""") | ChatOpenAI() | StrOutputParser()

# Add the routing logic - use the action key to route
def select_chain(output):
    if output["action"] == "SQL":
        return agent_executor
    elif output["action"] == "OTHER":
        return llm_chain
    else:
        raise ValueError

# Create the final chain
# Generate the action, and pass that into `select_chain`
chain = RunnableMap({
    "action": router_chain,
    "input": lambda x: x["question"]
})| select_chain

chain.invoke({"question":"How many SQL tables?"})

[csprocket777]

Thanks for this. I think I achieved something similar in JS/TS. I was putting too much responsibility onto LangChain for this.

[MANOJ21K]

@hwchase17 is there any way, lets say I have 50 tables in DB but filter top 5 relevant tables and then use its data for SQL agent..?

[AqeelMadhi]

@hwchase17 Hey, hope you can help me out here, how do i add memory to the RunnableMap and share the history between the two agents?