Allowing Multiple Outputs in Langchain's LLMChain

[CD-rajveer]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

1. Modify SequentialChain:

The SequentialChain class can be modified to handle situations where multiple outputs are desired from the chained LLM models. We can achieve this by changing the _run_output_key method.

With the modified SequentialChain

we can use new_chain.run and access the outputs as a dictionary.

results = new_chain.run("video of tutorial of how to make a chatbot for autocnfiguration of edge devices")

# Access outputs using dictionary keys
heading = results["heading"]
overview = results["overview"]

print(f"Youtube Title: {heading}")
print(f"Youtube Description: {overview}")

Motivation

Proposal (If applicable)

This proposal outlines the implementation of a new feature in Langchain's LLMChain class to handle scenarios where multiple outputs are desired from chained Large Language Models (LLMs).

Current Limitation:

The current LLMChain class is designed to work with a single output variable. This limitation restricts its use in situations where multiple outputs are necessary from the chained LLMs.

Proposed Solution:

We propose modifying the LLMChain class to accommodate multiple outputs. This can be achieved by:

Modifying _run_output_key method:
Check the number of output keys defined in the chain (self.output_keys).
If there's only one output key, use the existing behavior for running the chain and returning the single output.
If there are multiple outputs:
Run each chain within the sequence.
Create a dictionary to store the outputs.
Use the chain's name as the key and the chain's output as the value in the dictionary.
Return the dictionary containing all outputs.
Updating run method usage:
The run method of the sequential chain (SequentialChain) can be modified to accept this change.
When calling run on a chain with multiple outputs, the results will be returned as a dictionary.