SimpleSequentialChain Question

Short Course Q&A LangChain for LLM Application Development
1

For the SimpleSequentialChain example, the first part is:

llm = ChatOpenAI(temperature=0.9)
# prompt template 1
first_prompt = ChatPromptTemplate.from_template(
    "What is the best name to describe \
    a company that makes {product}?"
)
# chain 1
chain_one = LLMChain(llm=llm, prompt=first_prompt)

The second part is:

# prompt template 2
second_prompt = ChatPromptTemplate.from_template(
    "Write a 20 words description for the following \
    company:{company_name}"
)
# chain 2
chain_two = LLMChain(llm=llm, prompt=second_prompt)

Where is the result of the first prompt being assigned to the variable “{company_name}”?

1 Like
2

I have the same question; I have tested the 3rd prompt like this, without assigning {company_description}:

third_prompt = ChatPromptTemplate.from_template(
    "Write a 10 words slogan for the description:{company_description}"
)
chain_three = LLMChain(llm=llm, prompt=third_prompt)

And the result is:

Royal Bedding Co.
Royal Bedding Co. is a luxurious bedding company that provides high-quality bedding products for an exceptional sleeping experience.
“Indulge in luxury: Sleep better with Royal Bedding Co.”

My thought is that LLMs were already pre-trained with a quantity of codes, so they’ve learnt that we were in a “chain” and knew what we wanted to achieve. Based on this context, AI can do the following without us assigning anything.
Correct me I’m wrong. @gent.spah @ai_curious

3

BTW this post should be replaced under LangChain for LLM Application Development

4

What do you mean, I am not understanding it fully?

5

Ah huh, this is the background:

It’s in lesson 4 of LangChain for LLM Application Development. The SimpleSequentialChain code demo is as follows:

from langchain.chains import SimpleSequentialChain
llm = ChatOpenAI(temperature=0.9)

# prompt template 1
first_prompt = ChatPromptTemplate.from_template(
    "What is the best name to describe \
    a company that makes {product}?"
)

# Chain 1
chain_one = LLMChain(llm=llm, prompt=first_prompt)

# prompt template 2
second_prompt = ChatPromptTemplate.from_template(
    "Write a 20 words description for the following \
    company:{company_name}"
)
# chain 2
chain_two = LLMChain(llm=llm, prompt=second_prompt)

overall_simple_chain = SimpleSequentialChain(chains=[chain_one, chain_two],

overall_simple_chain.run(product)

The {product} in first_prompt is from overall_simple_chain.run(product).
But how about {company_name} in second_prompt ? It’s not assigned anywhere.

6

Yeah its counterintuitive but what I think is happening here is you are creating chains linked together chain_one = LLMChain(llm=llm, prompt=first_prompt) is linked with
chain_two = LLMChain(llm=llm, prompt=second_prompt)

and there is a memory of prompts here so there is no need to go through all the steps manually.

7

I agree with you; they’re already chained so the input of the next chain is the output of the previous chain by default

8

I think its because of LangChain here rather than the LLM itself.

9

I’m still looking for the source code to see the exact mechanism, but I can’t believe it is a coincidence that the first prompt in the chain asks the LLM to make up a name of a company, and the second prompt in the chain uses {company_name}. There are some hints about aligning output_variables and input_variables here: Sequential | 🦜️🔗 Langchain

  • SimpleSequentialChain: The simplest form of sequential chains, where each step has a singular input/output, and the output of one step is the input to the next.

my emphasis added

if you print out the overall_simple_chain object before calling run() you can see

output_key='text'

from the first prompt, immediately followed by

PromptTemplate(input_variables=['company_name']

for the second prompt. I assume that in the source code of SimpleSequentialChain we could see the actual mapping/assignment happening.

EDIT:

yeah, it ain’t rocket science…

https://github.com/hwchase17/langchain/blob/master/langchain/chains/sequential.py