Créer une chaîne d'extraction hybride

Graph RAG avec LangChain et Neo4j

Adam Cowley

Manager, Developer Education at Neo4j

Combiner les extractions

Rendre des fonctions exécutables

from langchain_core.runnables import RunnableLambda


double_chain = RunnableLambda(
    lambda input: input["input"] * 2
)


double_chain.invoke({"input": 2})

Transferts directs exécutables

from langchain_core.runnables import RunnablePassthrough


double_passthrough = RunnablePassthrough.assign(
    doubled=RunnableLambda(lambda x: x["input"] * 2)
)


double_passthrough.invoke({"input": 2})

{"input": 2, "doubled": 4}

Créer une invite de questions-réponses

graphrag_qa_prompt = ChatPromptTemplate.from_messages([

    SystemMessagePromptTemplate.from_template("""
    You are a helpful assistant answering questions about the play Romeo and Juliet.

    You are given a question and a context.

    Question: {input}
    """),

Créer une invite de questions-réponses

    ...
    SystemMessagePromptTemplate.from_template("""
    The following context has been retrieved from the database using vector search to help
    you answer the question: 

    {vectors}
    """),

Créer une invite de questions-réponses

    ...
    SystemMessagePromptTemplate.from_template("""
    The following data has been retrieved from the knowledge graph using Cypher to 
    answer the question.  

    {records}

    You can treat any information contained from the knowledge graph as authoritative.

    If the information does not exist in this answer, fall back to vector search results.

    If the answer is not included in either just say that you don't know and don't rely
    on pre-existing knowledge."""),

Créer une chaîne de questions-réponses

# line_retriever = Neo4jVector().as_retriever()
# text_to_cypher_chain = text_to_cypher_prompt | llm | StrOutputParser()
# graph = Neo4jGraph()


graphrag_qa_chain = RunnablePassthrough.assign(

    vectors=RunnableLambda(lambda x: line_retriever.invoke(x["input"])),

    records=text_to_cypher_chain | graph.query

)

    | graphrag_qa_prompt

    | llm

    | StrOutputParser()

Exécuter la chaîne de questions-réponses

graphrag_qa_chain.invoke({"input": "Who is Romeo's best friend?"})

Romeo's best friends are Mercutio and Benvolio.

Passons à la pratique !

Graph RAG avec LangChain et Neo4j