Chaining two prompts while keeping memory

[kevin-guimard-ext]

I want to chain two prompts of this kind:

"Give a list of 10 {X} players"
... => waiting for IA's ranswer
"For every player, give their birth date"

X being the name of a sport.

I've implemented this with 2 independent chains and linked both with a SimpleSequentialChain.

Is it the right process? When I create the overall chain, I get the following error message:
Chains used in SimplePipeline should all have one input
The fact is my second chain does not have inputs in the prompt since the inputs are given by the IA's answer. How can I handle such a chain? Do I have to use a ConversationBufferMemory class? I don't find the examples given in the documentation very insightful...

[dosubot[bot]]

Answer generated by a 🤖

Answer

Hello,

From your description, it seems like you're trying to chain two prompts where the second chain's input is dependent on the first chain's output. The error message you're encountering suggests that one of your chains is expecting more than one input.

In LangChain, the SimpleSequentialChain is designed to handle chains where each chain has exactly one input and one output. If a chain does not meet these criteria, it raises a ValueError. This might be the reason you're encountering the error.

To resolve this issue, you need to modify your second chain to only expect one input, which should be the output of the first chain. Here's an example of how you can modify your second chain:

combine_prompt_template = """Given the following extracted parts of a long document and a question, create a final answer with references ("SOURCES"). 
If you don't know the answer, just say that you don't know. Don't try to make up an answer.
ALWAYS return a "SOURCES" part in your answer.
{input}
=========
FINAL ANSWER:"""
COMBINE_PROMPT = PromptTemplate(
    template=combine_prompt_template, input_variables=["input"]
)

In this modification, the second chain now expects a single input, which should be the output of the first chain. This way, you can chain the two prompts together without encountering the error.

As for maintaining context between chains, the ConversationBufferMemory class in LangChain can be used. It provides a buffer for storing the conversation history and can be used to maintain context between chains by storing the history of the conversation and loading it when needed.

I hope this helps! If you have any other questions or need further clarification, feel free to ask.

Best,
Dosu

Sources

• libs/langchain/langchain/chains/qa_with_sources/map_reduce_prompt.py
• libs/langchain/langchain/chains/sequential.py
• docs/snippets/modules/chains/foundational/sequential_chains.mdx
• libs/langchain/langchain/memory/buffer.py
• libs/langchain/langchain/chains/conversation/base.py

---

This response is meant to be useful, save you time, and share context. 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.

[kevin-guimard-ext]

Hello,

Thanks for your quick reply. Now, as you've seen in my example, no input appears in the second prompt so I cannot use your solution. In fact, I want to model a conversation made of a call to GPT, an answer, and a second call to GPT while keeping the previous context. So, I tried the ConversationBufferMemory class:

chat_model = AzureChatOpenAI(...)

prompt = ChatPromptTemplate.from_messages([
    HumanMessagePromptTemplate(prompt=PromptTemplate(
           template="Give a list of 10 {sport} players.",
           input_variables=["sport"])),
    MessagesPlaceholder(variable_name="chat_history"),
    HumanMessage(content="For every player, give their birth date."),
])

memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True)

chain = LLMChain(
    llm=chat_model,
    prompt=prompt,
    verbose=True,
    memory=memory,
)
chain.run(sport="soccer")

The result is correct but the logs seem to show that the 2 prompts are chained before the call to GPT, which is not what I expect. I'd like my chain to trigger 2 calls to GPT: one after the first prompt and one after the other. Is it clear?

[dosubot[bot]]

Answer generated by a 🤖

Answer

Hello,

From your description, it seems like you want to make two separate calls to the GPT model, one after each prompt. However, the current implementation of LLMChain in LangChain chains the prompts before making a call to the language model.

To achieve what you want, you would need to modify your code to call the format method on each PromptTemplate separately and then pass the result to the language model. Here's an example of how you can modify your code:

from langchain import PromptTemplate

# Create two separate PromptTemplates
prompt1 = PromptTemplate.from_template("Say {foo}")
prompt2 = PromptTemplate.from_template("Then say {bar}")

# Format each prompt separately
formatted_prompt1 = prompt1.format(foo="Hello")
formatted_prompt2 = prompt2.format(bar="World")

# Pass each formatted prompt to GPT separately
gpt_response1 = gpt_model.generate(formatted_prompt1)
gpt_response2 = gpt_model.generate(formatted_prompt2)

In this example, gpt_model is a hypothetical instance of your GPT model. You would replace this with the actual code to call your GPT model. The key point is that you're calling gpt_model.generate (or whatever method you use to call GPT) separately for each formatted prompt.

Please note that this approach will require you to manage the conversation history manually if you want the second prompt to be aware of the context from the first prompt's output.

I hope this helps! If you have any further questions, feel free to ask.

Sources

• libs/langchain/langchain/prompts/prompt.py
• libs/langchain/langchain/chains/llm.py

---

This response is meant to be useful, save you time, and share context. 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.