Améliorer la recherche dans le graphe

Retrieval Augmented Generation (RAG) avec LangChain

Meri Nova

Machine Learning Engineer

Techniques

Limitation principale : fiabilité de la traduction utilisateur → Cypher

Stratégies pour améliorer le système de recherche dans le graphe :

Filtrer le schéma du graphe
Valider la requête Cypher
Few-shot prompting

Filtrage

from langchain_community.chains.graph_qa.cypher import GraphCypherQAChain

llm = ChatOpenAI(api_key="...", model="gpt-4o-mini", temperature=0)


chain = GraphCypherQAChain.from_llm(

    graph=graph, llm=llm, exclude_types=["Concept"], verbose=True

)

print(graph.get_schema)

Propriétés des nœuds :
Document {title: STRING, id: STRING, text: STRING, summary: STRING, source: STRING}
Organization {id: STRING}

Validation de la requête Cypher

Difficulté à interpréter la direction des relations

chain = GraphCypherQAChain.from_llm(
    graph=graph, llm=llm, verbose=True, validate_cypher=True
)

Détecte les nœuds et les relations
Détermine la direction des relations
Vérifie le schéma du graphe
Met à jour la direction des relations

Few-shot prompting

examples = [
    {
        "question": "How many notable large language models are mentioned in the article?",
        "query": "MATCH (m:Concept {id: 'Large Language Model'}) RETURN count(DISTINCT m)",
    },
    {
        "question": "Which companies or organizations have developed the large language models mentioned?",
        "query": "MATCH (o:Organization)-[:DEVELOPS]->(m:Concept {id: 'Large Language Model'}) RETURN DISTINCT o.id",
    },
    {
        "question": "What is the largest model size mentioned in the article, in terms of number of parameters?",
        "query": "MATCH (m:Concept {id: 'Large Language Model'}) RETURN max(m.parameters) AS largest_model",
    },
]

Mettre en œuvre le few-shot prompting

from langchain_core.prompts import FewShotPromptTemplate, PromptTemplate


example_prompt = PromptTemplate.from_template(
    "User input: {question}\nCypher query: {query}"
)


cypher_prompt = FewShotPromptTemplate(
    examples=examples,
    example_prompt=example_prompt,
    prefix="You are a Neo4j expert. Given an input question, create a syntactically correct
    Cypher query to run.\n\nHere is the schema information\n{schema}.\n\n
    Below are a number of examples of questions and their corresponding Cypher queries.",
    suffix="User input: {question}\nCypher query: ",
    input_variables=["question"],
)

Invite complète

You are a Neo4j expert. Given an input question, create a syntactically correct Cypher query to run.

Below are a number of examples of questions and their corresponding Cypher queries.

User input: How many notable large language models are mentioned in the article?
Cypher query: MATCH (p:Paper) RETURN count(DISTINCT p)

User input: Which companies or organizations have developed the large language models?
Cypher query: MATCH (o:Organization)-[:DEVELOPS]->(m:Concept {id: 'Large Language Model'}) RETURN DISTINCT o.id

User input: What is the largest model size mentioned in the article, in terms of number of parameters?
Cypher query: MATCH (m:Concept {id: 'Large Language Model'}) RETURN max(m.parameters) AS largest_model

User input: How many papers were published in 2016?
Cypher query:

Ajouter les exemples few-shot

chain = GraphCypherQAChain.from_llm(
    graph=graph, llm=llm, cypher_prompt=cypher_prompt,
    verbose=True, validate_cypher=True
)

Passons à la pratique !

Retrieval Augmented Generation (RAG) avec LangChain