Memory graphs

Graph RAG with LangChain and Neo4j

Adam Cowley

Manager, Developer Education at Neo4j

Memory in RAG applications

Short-term memory

Short-lived, scoped lifespan
Conversation history

Long-term memory

Semantic memory: facts about the world
Useful for hyper-personalization

Short-term memory

Short-term to long-term memory

Neo4j chat message history

from langchain_neo4j import Neo4jChatMessageHistory


history = Neo4jChatMessageHistory(

    url=NEO4J_URI,

    username=NEO4J_USERNAME,

    password=NEO4J_PASSWORD,

    session_id="session_id_1",

)


# Add a human message
history.add_user_message("hi!")


# Add an AI message
history.add_ai_message("what's up?")

Problems with storing everything

Neo4j chat message history

from langchain_neo4j import Neo4jChatMessageHistory

history = Neo4jChatMessageHistory(
    url=NEO4J_URI,
    username=NEO4J_USERNAME,
    password=NEO4J_PASSWORD,
    session_id="session_id_1",

    window=10 # defaults to 3 messages

)


# Get history
print(history.messages)

[HumanMessage(content='hi!', ...), AIMessage(content='whats up?', ...)]

Summarizing conversations

from pydantic import BaseModel, Field

class ConversationFact(BaseModel):
    """
    A class that holds the facts from a conversation in a format of object, subject, predicate.
    """

    object: str = Field(description="The object of the fact. For example, 'Adam' ")

    subject: str = Field(description="The subject of the fact. For example, 'Ice cream'")

    relationship: str = Field(description="The relationship between object and subject. Eg: 'LOVES'")


class ConversationFacts(BaseModel):
    """
    A class that holds a list of ConversationFact objects.
    """
    facts: list[ConversationFact] = Field(description="A list of ConversationFact objects.")

Summarizing conversations

llm_with_output = (
    ChatOpenAI(model="gpt-4o-mini", openai_api_key=OPENAI_API_KEY)
        .with_structured_output(ConversationFacts)
)

Summarizing conversations

prompt = ChatPromptTemplate.from_messages(

    SystemMessagePromptTemplate.from_template("Extract the facts from the conversation."),

    MessagesPlaceholder(variable_name="history"),

)


chain = prompt | llm_with_output


chain.invoke({

    "history": history.messages,

})

ConversationFacts(facts=[
    ConversationFact(object='child', subject='Bluey', relationship='LOVES')
])

Let's practice!

Graph RAG with LangChain and Neo4j