Multipromtchain with RetrievalQA as destination chain

[DivyaNautiyal-Rev]

I have multiple RetrievalQA chains with different format instructions and prompt template, I am using MultiPromptChain from langchain to route it to the suitable chain, however I am facing error with the inputs, I tried changing the variable as 'input' in my prompt template but it doesn't help.

Here is the error that I am facing :
ValueError: Missing some input keys: {'query'}

Here is the implementation of my MultiPromptChain
`class CustomMultiPromptChain (MultiRouteChain):

destination_chains: Mapping[str, Chain]
"""Map of name to candidate chains that inputs can be routed to. Not restricted to LLM"""

`

and here's how I am initializing and using it :
`destination_chains = {}
for p_info in prompt_infos:
name = p_info["name"]
# prompt = p_info["prompt"]
chain_kwargs = p_info["kwargs"]
chain = RetrievalQA.from_chain_type(llm=ChatOpenAI(model = "gpt-4"), chain_type="stuff", retriever=retriever,
chain_type_kwargs=chain_kwargs)
destination_chains[name] = chain

chain = CustomMultiPromptChain(
    router_chain=router_chain,
    destination_chains=destination_chains,
    default_chain=default_chain,
    verbose=True,
)

chain({"input":input})

`
Any kind of help/suggestion is appreciated.

[dosubot[bot]]

🤖

Hello,

The error you're encountering is due to the fact that the MultiPromptChain is expecting a query key in the input, but it's not being provided. The RetrievalQA chain's prompt template, as shown in your context, requires two input variables: context and question.

To resolve this issue, you need to ensure that the input you're providing to the MultiPromptChain includes a query key. The query should be a dictionary that includes both context and question keys. Here's an example of how you might structure your input:

input_data = {
    'query': {
        'context': 'Your context here',
        'question': 'Your question here'
    }
}

Then, you can pass this input_data to your MultiPromptChain:

result = multi_prompt_chain(input_data)

This should resolve the ValueError: Missing some input keys: {'query'} issue you're encountering.

Additionally, if you have a situation where one or more destination chains are expecting a different input variable, you can create a custom chain that adapts the input variables for the destination chain. This custom chain will take the input variable from the router chain and convert it to the expected input variable for the destination chain.

Here's an example of how you can create such a custom chain:

from langchain.chains.base import Chain

class InputAdapterChain(Chain):
    def __init__(self, destination_chain, input_key_map):
        self.destination_chain = destination_chain
        self.input_key_map = input_key_map

    def run(self, inputs):
        adapted_inputs = {self.input_key_map[k]: v for k, v in inputs.items()}
        return self.destination_chain.run(adapted_inputs)

Now, you can use this InputAdapterChain to wrap the PalChain and adapt the input variable:

pal_chain = PalChain()  # Assuming you have initialized the PalChain
input_key_map = {"input": "question"}
adapted_pal_chain = InputAdapterChain(pal_chain, input_key_map)

Finally, use the adapted_pal_chain in your MultiPromptChain setup:

chain = MultiPromptChain(router_chain=router_chain, 
                         destination_chains=[adapted_pal_chain], 
                         default_chain=default_chain, 
                         verbose=True
                        )

This way, the InputAdapterChain will convert the input variable from the router chain to the expected input variable for the PalChain, avoiding the ValueError you encountered.

I hope this helps! Let me know if you have any other questions.

Sources

• Issue: How to handle RouterChain when 1 or more destination chain(s) which is expecting a different input variable?
• Issue: Missing some input keys: {'question'} when using ConversationalRetrievalChain
• libs/langchain/langchain/chains/retrieval_qa/prompt.py
• libs/langchain/langchain/chains/router/multi_prompt.py

---

This response is meant to be useful and save you time. It is not meant to be a precise solution, but rather a starting point for your own research.

Help me be more useful! Please leave a 👍 if this is helpful and 👎 if it is irrelevant.

If you want to continue the conversation, start your reply with @dosu-bot.

[DivyaNautiyal-Rev]

@dosu-bot
As far as I am aware context is determined itself by the retriever in RetrievalQAChain and I cannot pass it externally, how would I find out context?

[dosubot[bot]]

🤖

Hello,

The error you're encountering, ValueError: Missing some input keys: {'query'}, is likely due to the fact that the MultiPromptChain is expecting a query input, but it's not being provided.

In the context shared, the MultiRetrievalQAChain class is used, which is a multi-route chain that uses an LLM router chain to choose amongst retrieval QA chains. The from_retrievers class method is used to create an instance of this class. This method creates a router_chain using the LLMRouterChain.from_llm method, which expects a query input as specified by the RouterOutputParser(next_inputs_inner_key="query").

To resolve this issue, you need to ensure that the query input is being provided to the MultiRetrievalQAChain. If the query is determined by the retriever in the RetrievalQAChain, you might need to modify the RetrievalQAChain or the MultiRetrievalQAChain to pass the query along to the next chain in the sequence.

Unfortunately, without the actual code of your MultiRetrievalQAChain and RetrievalQAChain setup, it's hard to provide a more specific solution. But generally, you need to ensure that all expected inputs for each chain in your MultiRetrievalQAChain are being provided.

Additionally, there was a similar issue in the LangChain repository: Issue: How to handle RouterChain when 1 or more destination chain(s) which is expecting a different input variable?. The solution suggested there was to create a custom chain that adapts the input variables for the destination chain. This custom chain takes the input variable from the router chain and converts it to the expected input variable for the destination chain. Here's an example of how you can create such a custom chain:

from langchain.chains.base import Chain

class InputAdapterChain(Chain):
    def __init__(self, destination_chain, input_key_map):
        self.destination_chain = destination_chain
        self.input_key_map = input_key_map

    def run(self, inputs):
        adapted_inputs = {self.input_key_map[k]: v for k, v in inputs.items()}
        return self.destination_chain.run(adapted_inputs)

Now, you can use this InputAdapterChain to wrap the RetrievalQA chain and adapt the input variable:

input_key_map = {"input": "question"}
adapted_chain = InputAdapterChain(chain, input_key_map)

Finally, use the adapted_chain in your MultiPromptChain setup:

chain = CustomMultiPromptChain(
    router_chain=router_chain,
    destination_chains={name: adapted_chain for name, chain in destination_chains.items()},
    default_chain=default_chain,
    verbose=True,
)

This way, the InputAdapterChain will convert the input variable from the router chain to the expected input variable for the RetrievalQA chain, avoiding the ValueError you encountered.

I hope this helps! Let me know if you have any other questions.

Sources

• Issue: How to handle RouterChain when 1 or more destination chain(s) which is expecting a different input variable?
• libs/langchain/langchain/chains/retrieval_qa/prompt.py
• libs/langchain/langchain/chains/router/multi_retrieval_qa.py
• libs/langchain/tests/integration_tests/chains/test_retrieval_qa.py
• libs/langchain/langchain/chains/retrieval_qa/base.py

---

This response is meant to be useful and save you time. It is not meant to be a precise solution, but rather a starting point for your own research.

Help me be more useful! Please leave a 👍 if this is helpful and 👎 if it is irrelevant.

If you want to continue the conversation, start your reply with @dosu-bot.