How to build a reusable chain that takes the output of another chain as a keyword argument for the template?

[adrenaline681]

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

# Title Chain
title_template = PromptTemplate.from_template('Generate a title for a blog post about the following topic: {topic}')
generate_title = title_template | llm | StrOutputParser()

# Translate Chain
translate_template = PromptTemplate.from_template('Translate the following text to {language}: {text}')
translate_chain = {'language': itemgetter('language'), 'text': HOW_TO_PASS_OUTPUT_OF_PREVIOUS_CHAIN?} | translate_template | llm | StrOutputParser()

# Combine and run
combined_chain = generate_title | translate_chain
result = combined_chain.invoke({'topic': 'The benefits of exercise', 'language': 'Spanish'})

Description

I'm trying to build reusable chains and I'm having issues trying to set it up so it takes the output of the previous chain as one of the template format arguments.

In this example, I have a chain that generates a title based on a topic and a chain that translates the title into a different language. The topic and the language I want to provide them when I invoke the final chain.

If my 2nd chain only has one single input for the template (because I hardcode the language), everything works perfectly:

# Title Chain
title_template = PromptTemplate.from_template('Generate a title for a blog post about the following topic: {topic}')
generate_title = title_template | llm | StrOutputParser()

# Translate Chain
translate_template = PromptTemplate.from_template('Translate the following text to spanish: {text}')
translate_chain = translate_template | llm | StrOutputParser()

# Combine and run
combined_chain = generate_title | translate_chain
result = combined_chain.invoke({'topic': 'The benefits of exercise'})

But if I want to be able to provide the language when I invoke the final chain then, I dont know how to pass the output to the template dictionary. See the "HOW_TO_PASS_OUTPUT_OF_PREVIOUS_CHAIN" in the code

# Title Chain
title_template = PromptTemplate.from_template('Generate a title for a blog post about the following topic: {topic}')
generate_title = title_template | llm | StrOutputParser()

# Translate Chain
translate_template = PromptTemplate.from_template('Translate the following text to {language}: {text}')
translate_chain = {'language': itemgetter('language'), 'text': HOW_TO_PASS_OUTPUT_OF_PREVIOUS_CHAIN?} | translate_template | llm | StrOutputParser()

# Combine and run
combined_chain = generate_title | translate_chain
result = combined_chain.invoke({'topic': 'The benefits of exercise', 'language': 'Spanish'})

I also tried to use a lambda function but the itemgetter doesn't seem to work:

translate_chain = (
    lambda text: {'language': itemgetter('language'), 'text': text}
    | translate_template
    | llm
    | StrOutputParser()
)

If anyone knows the answer I would be much appreciated.

System Info

langchain==0.2.6
langchain-community==0.2.6
langchain-core==0.2.10
langchain-openai==0.1.12
langchain-text-splitters==0.2.2

Windows

Python 3.12

[dosubot[bot]]

Hey there, @adrenaline681! I'm Dosu, a bot here to assist you with any bugs, questions, or contributions you have in mind. Let's tackle this challenge together while we wait for a human maintainer.

To build a reusable chain that takes the output of another chain as a keyword argument for the template, you can use the RunnableParallel class to combine chains and then pass the output as needed. Here is an example based on your provided code:

from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import PromptTemplate
from langchain_core.runnables import RunnableParallel, RunnableLambda
from langchain_openai import OpenAI

# Initialize the language model
llm = OpenAI(model="text-davinci-003", temperature=0)

# Define the title generation prompt template
title_template = PromptTemplate.from_template('Generate a title for a blog post about the following topic: {topic}')
generate_title = title_template | llm | StrOutputParser()

# Define the translation prompt template
translate_template = PromptTemplate.from_template('Translate the following text to {language}: {text}')

# Create a custom function to handle the output
def translate_text(inputs):
    text = inputs['text']
    language = inputs['language']
    return translate_template.invoke({'text': text, 'language': language})

# Wrap the custom function in a RunnableLambda
translate_text_runnable = RunnableLambda(translate_text)

# Create the composed chain
composed_chain = (
    RunnableParallel({"text": generate_title})
    .pipe(translate_text_runnable)
    .pipe(llm)
    .pipe(StrOutputParser())
)

# Invoke the composed chain
result = composed_chain.invoke({'topic': 'The benefits of exercise', 'language': 'Spanish'})
print(result)

In this example:

1. The first chain generates a title based on the topic.
2. The RunnableParallel class is used to pass the output of the first chain as the text keyword argument to the translate_template.
3. A custom function translate_text is created to handle the output and wrapped in a RunnableLambda.
4. The composed chain then translates the generated title into the specified language using the translate_template, the model, and the StrOutputParser.

This approach ensures that the output of one chain is correctly passed as a keyword argument to another template in a reusable manner ^[1][2][3].

To continue talking to Dosu, mention @dosu.

[adrenaline681]

This solution doesnt seem to work when I run it.

Isnt there an easy way to specify in the dictionary where you want to place the output of the previous step? Something like this:

translate_chain = {'language': itemgetter('language'), 'text': PreviousStepOutput()} | translate_template | llm | StrOutputParser()

[dosubot[bot]]

To pass the output of one chain as a keyword argument to another chain's template in LangChain, you can use the pipe operator (|) to chain runnables together more seamlessly. Here is an example based on your provided code:

from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import PromptTemplate
from langchain_openai import OpenAI

# Initialize the language model
llm = OpenAI(model="text-davinci-003", temperature=0)

# Define the title generation prompt template
title_template = PromptTemplate.from_template('Generate a title for a blog post about the following topic: {topic}')
generate_title = title_template | llm | StrOutputParser()

# Define the translation prompt template
translate_template = PromptTemplate.from_template('Translate the following text to {language}: {text}')

# Create the translation chain
translate_chain = translate_template | llm | StrOutputParser()

# Combine the chains
combined_chain = generate_title | (lambda output: {"text": output, "language": itemgetter('language')}) | translate_chain

# Invoke the combined chain
result = combined_chain.invoke({'topic': 'The benefits of exercise', 'language': 'Spanish'})
print(result)

In this example:

1. The generate_title chain generates a title based on the topic.
2. The output of generate_title is passed to a lambda function that constructs the input dictionary for the translate_chain.
3. The translate_chain then translates the generated title into the specified language.

This approach allows you to specify where you want to place the output of the previous step directly in the dictionary ^[1][2].

To continue talking to Dosu, mention @dosu.