Passing instructions to LLM via Prompt-Question

[mraguth]

I prefer to provide clear instructions to the LLM through the prompt. My question is whether I can add new lines to the 'query_transform_prompt' or if I should create a new prompt and reference it in the code. This is my code—please help

from langchain_core.output_parsers import StrOutputParser
from langchain_core.runnables import RunnableBranch

# We need a prompt that we can pass into an LLM to generate a transformed search query

chat = ChatOpenAI(model="gpt-4o-mini", temperature=0.2)

query_transform_prompt = ChatPromptTemplate.from_messages(
    [
        MessagesPlaceholder(variable_name="messages"),
        (
            "user",
            """Given the above conversation, generate a search query to look up in order to get information relevant to the conversation. Only respond with the query, nothing else.,
            Write appropriate context for each and every task based on what you seen in the query. In case conversation has no content related to query, please feel free to come up with real life expereince""",
        ),
    ]
)

query_transforming_retriever_chain = RunnableBranch(
    (
        lambda x: len(x.get("messages", [])) == 1,
        # If only one message, then we just pass that message's content to retriever
        (lambda x: x["messages"][-1].content) | retriever,
    ),
    # If messages, then we pass inputs to LLM chain to transform the query, then pass to retriever
    query_transform_prompt | chat | StrOutputParser() | retriever,
).with_config(run_name="chat_retriever_chain")

document_chain = create_stuff_documents_chain(chat, question_answering_prompt)

conversational_retrieval_chain = RunnablePassthrough.assign(
    context=query_transforming_retriever_chain,
).assign(
    answer=document_chain,
)

demo_ephemeral_chat_history = ChatMessageHistory()

demo_ephemeral_chat_history.add_user_message(query)                 ##### This is the input

response = conversational_retrieval_chain.invoke(
    {"messages": demo_ephemeral_chat_history.messages},
)

demo_ephemeral_chat_history.add_ai_message(response["answer"])

print(str(response))

[dosubot[bot]]

You can add new lines to the query_transform_prompt by simply formatting the string within the prompt template. Python's triple-quoted strings ("""...""") allow you to include new lines directly in the string. Here's how you can modify your existing prompt to include new lines for better readability:

query_transform_prompt = ChatPromptTemplate.from_messages(
    [
        MessagesPlaceholder(variable_name="messages"),
        (
            "user",
            """Given the above conversation, generate a search query to look up in order to get information relevant to the conversation. Only respond with the query, nothing else.
            Write appropriate context for each and every task based on what you see in the query. In case the conversation has no content related to the query, please feel free to come up with a real-life experience."""
        ),
    ]
)

This approach keeps your code clean and maintains the prompt's readability without needing to create a new prompt object. Just ensure that the logic and structure of your prompt remain consistent with your intended instructions for the LLM.

To continue talking to Dosu, mention @dosu.

---

^{Help Dosu learn! Give it feedback: Great Response | Irrelevant Answer | Incorrect Sources | Too Verbose | Hallucination | Other}

[mraguth]

@dosu. Thank you. If I like to pass few shot prompt template? how I do that? Below is my sample query.

# Create the few-shot prompt template using the examples and example prompt
few_shot_prompt_rec = FewShotChatMessagePromptTemplate(
    examples=examples,
    example_prompt=example_prompt,
)

[dosubot[bot]]

To pass a few-shot prompt template using the FewShotChatMessagePromptTemplate, you can follow this example:

from langchain_core.prompts import ChatPromptTemplate, FewShotChatMessagePromptTemplate
from langchain_core.prompts import HumanMessagePromptTemplate, AIMessagePromptTemplate

# Define examples
examples = [
    {"input": "2+2", "output": "4"},
    {"input": "2+3", "output": "5"},
]

# Create an example prompt using ChatPromptTemplate
example_prompt = ChatPromptTemplate.from_messages(
    [
        HumanMessagePromptTemplate.from_template("{input}"),
        AIMessagePromptTemplate.from_template("{output}"),
    ]
)

# Create a FewShotChatMessagePromptTemplate
few_shot_prompt = FewShotChatMessagePromptTemplate(
    input_variables=["input"],
    example_prompt=example_prompt,
    examples=examples,
)

# Combine with a system message and a human message
final_prompt = ChatPromptTemplate.from_messages(
    [
        ("system", "You are a helpful AI Assistant"),
        few_shot_prompt,
        ("human", "{input}"),
    ]
)

# Format the messages with a new input
messages = final_prompt.format_messages(input="100 + 1")

This code sets up a few-shot prompt template with examples and combines it with a system message to create a final prompt. You can then use the format_messages method to generate the messages with a new input ^[1].

To continue talking to Dosu, mention @dosu.

---

^{Help Dosu learn! Give it feedback: Great Response | Irrelevant Answer | Incorrect Sources | Too Verbose | Hallucination | Other}

[mraguth]

@dosu... here how I add 'few_shot_prompt' in the below code?

query_transforming_retriever_chain = RunnableBranch(
    (
        lambda x: len(x.get("messages", [])) == 1,
        # If only one message, then we just pass that message's content to retriever
        (lambda x: x["messages"][-1].content) | retriever,
    ),
    # If messages, then we pass inputs to LLM chain to transform the query, then pass to retriever
    query_transform_prompt | chat | StrOutputParser() | retriever,
).with_config(run_name="chat_retriever_chain")

document_chain = create_stuff_documents_chain(chat, question_answering_prompt)

conversational_retrieval_chain = RunnablePassthrough.assign(
    context=query_transforming_retriever_chain,
).assign(
    answer=document_chain,
)

demo_ephemeral_chat_history = ChatMessageHistory()

demo_ephemeral_chat_history.add_user_message(query)                 ##### This is the input

response = conversational_retrieval_chain.invoke(
    {"messages": demo_ephemeral_chat_history.messages},
)

demo_ephemeral_chat_history.add_ai_message(response["answer"])

print(str(response))

[dosubot[bot]]

To integrate the few_shot_prompt into your existing code, you can replace the query_transform_prompt with the few_shot_prompt in the RunnableBranch. Here's how you can modify your code:

1. Ensure you have defined the few_shot_prompt as shown in the previous example.
2. Replace the query_transform_prompt with few_shot_prompt in the RunnableBranch.

Here's how the updated code would look:

# Assuming few_shot_prompt is already defined as shown in the previous example

query_transforming_retriever_chain = RunnableBranch(
    (
        lambda x: len(x.get("messages", [])) == 1,
        # If only one message, then we just pass that message's content to retriever
        (lambda x: x["messages"][-1].content) | retriever,
    ),
    # If messages, then we pass inputs to LLM chain to transform the query, then pass to retriever
    few_shot_prompt | chat | StrOutputParser() | retriever,
).with_config(run_name="chat_retriever_chain")

document_chain = create_stuff_documents_chain(chat, question_answering_prompt)

conversational_retrieval_chain = RunnablePassthrough.assign(
    context=query_transforming_retriever_chain,
).assign(
    answer=document_chain,
)

demo_ephemeral_chat_history = ChatMessageHistory()

demo_ephemeral_chat_history.add_user_message(query)  # This is the input

response = conversational_retrieval_chain.invoke(
    {"messages": demo_ephemeral_chat_history.messages},
)

demo_ephemeral_chat_history.add_ai_message(response["answer"])

print(str(response))

This modification uses the few_shot_prompt to format the input messages before passing them to the language model and the retriever. Make sure that the few_shot_prompt is correctly set up with examples and an example prompt as shown in the previous example.

To continue talking to Dosu, mention @dosu.

---

^{Help Dosu learn! Give it feedback: Great Response | Irrelevant Answer | Incorrect Sources | Too Verbose | Hallucination | Other}