runnable.py

from langchain.chains.openai_functions import create_structured_output_runnable
from langchain.chat_models import ChatOpenAI
from langchain.prompts import ChatPromptTemplate
from langchain.pydantic_v1 import BaseModel, Field
 
class Insight(BaseModel):
    insight: str = Field(description="""insight""")

chat_model = ChatOpenAI(model_name="gpt-4-1106-preview")

from langchain.prompts import MessagesPlaceholder

prompt = ChatPromptTemplate.from_messages(
    [
        (
            "system",
            """
            long system message
            """
        ),
        # This is the placeholder for chat history
        # We will add messages in the history of the conversation here over time
        MessagesPlaceholder(variable_name="history"),
        (
            "human",
            "What's the step with the biggest impact? ```{input}```",
        ),
        (
            "human",
            "Tip: Make sure to answer in the correct format"
        ),
    ]
)
 
runnable = create_structured_output_runnable(Insight, chat_model, prompt)
# We can call this without any history, but we have to pass in the key
response = runnable.invoke({"input": "Step 1: sign up\n\nStep 2: Do all the work", "history": []})

response

# We can also manually pass in chat history if we want
from langchain.schema.messages import AIMessage, HumanMessage

messages = [HumanMessage(content="hi!"), AIMessage(content="hi!")]

response = runnable.invoke({"input": "Step 1: sign up\n\nStep 2: Do all the work", "history": messages})

# Alternatively, we can use chat history to accumulate that automatically
# We will use an in memory implementation

# This is the in memory storage
from langchain.memory.chat_message_histories import ChatMessageHistory
# This is the runnable that will basically save inputs/outputs to memory
from langchain.schema.runnable.history import RunnableWithMessageHistory

# We need to initialize this upfront
memory = ChatMessageHistory()

chain_with_history = RunnableWithMessageHistory(
    # We need to add this extra `(lambda x: x.json())` to the end
    # That is because previously the return type was a Pydantic model - we don't know how to save that in history
    # In order to save in history, needs to be either a string or a Message type
    # If we wanted to do something different, we could! As long as its a string or message
    runnable | (lambda x: x.json()),
    # Since this is just in memory, session_id doesn't matter
    lambda session_id: memory,
    input_messages_key="input",
    # We need this to align with the `variable_name` in the prompt
    history_messages_key="history"
)

chain_with_history.invoke(
    {"input": "Step 1: sign up\n\nStep 2: Do all the work"},
    # Even though this isn't used by the chat message history, we still need to pass it in
    {'configurable': {'session_id': '123'}}
)

memory.messages