Ajouter la mémoire et la conversation

Concevoir des systèmes agentiques avec LangChain

Dilini K. Sumanapala, PhD

Founder & AI Engineer, Genverv, Ltd.

Tester l'usage des outils

Schéma complet du clavardeur avec le nœud d'outils ajouté.

Tester l'usage des outils

# Produce the chatbot graph
display(Image(app.get_graph().draw_mermaid_png()))


# Define a function to execute the chatbot, streaming each message
def stream_tool_responses(user_input: str):
    for event in graph.stream({"messages": [("user", user_input)]}):


        # Return the agent's last response
        for value in event.values():
            print("Agent:", value["messages"])


# Define the query and run the chatbot
user_query = "House of Lords"
stream_tool_responses(user_query)

Visualiser le diagramme

Diagramme du clavardeur LangGraph avec nœud d'outil

Diffuser le résultat en continu

Agent: [AIMessage(content='', additional_kwargs={'tool_calls': [{'function': {'arguments': 
'{"query":"House of Lords"}', 'name': 'wikipedia'}, 'type': 'function'}]},
response_metadata={'...'}])]

Agent: [ToolMessage(content='Page: House of Lords\nSummary: The House of Lords is the upper 
house of the Parliament of the United Kingdom. Like the lower house...The House of Lords 
also has a Church of England role...', name='wikipedia', id='...',')]

Agent: [AIMessage(content='The House of Lords is the upper house of the Parliament 
of the United Kingdom, located in the Palace of Westminster in London. It is one 
of the oldest institutions in the world..., additional_kwargs={},

response_metadata=
{'finish_reason': 'stop', 'model_name': 'gpt-4o-mini-2024-07-18', 'system_fingerprint': 
'fp_0ba0d124f1'}, id='run-ae3a0b4f-5b42-4f4e-9409-7c191af0b9c9-0')]

Ajouter la mémoire

# Import the modules for saving memory
from langgraph.checkpoint.memory import MemorySaver


# Modify the graph with memory checkpointing
memory = MemorySaver()


# Compile the graph passing in memory
graph = graph_builder.compile(checkpointer=memory)

Diffusion continue avec mémoire

# Set up a streaming function for a single user
def stream_memory_responses(user_input: str):

    config = {"configurable": {"thread_id": "single_session_memory"}}


    # Stream the events in the graph
    for event in graph.stream({"messages": [("user", user_input)]}, config):


        # Return the agent's last response
        for value in event.values():
            if "messages" in value and value["messages"]:
                print("Agent:", value["messages"])


stream_memory_responses("What is the Colosseum?")
stream_memory_responses("Who built it?")

Générer un résultat avec mémoire

stream_memory_responses("What is the Colosseum?")

Agent: [AIMessage(content='', additional_kwargs={'tool_calls': [{'index': 0, 
'id': '...', 'function': {'arguments': '
{"query":"Colosseum"}', 'name': 'wikipedia'}, ...])]

Agent: [ToolMessage(content='Page: Colosseum\nSummary: The Colosseum is an ancient
amphitheatre in Rome, Italy. It is the largest standing amphitheatre 
in the world..)]

Agent: [AIMessage(content='The Colosseum, located in Rome, is the largest ancient 
amphitheatre still standing. Built under Emperor Vespasian and completed by his son 
Titus, it hosted gladiatorial games and public events. It is also known as the Flavian 
Amphitheatre due to its association with the Flavian dynasty.', additional_kwargs={}, 
response_metadata={'finish_reason': 'stop', 'model_name': 'gpt-4o-mini-...', ...')]

Générer un résultat avec mémoire

stream_memory_responses("Who built it?")

Agent: [AIMessage(content='The Colosseum was built by Emperor Vespasian around 
72 AD and completed by his successor, Emperor Titus. Later modifications were 
made by Emperor Domitian...')]

Passons à la pratique !

Concevoir des systèmes agentiques avec LangChain