Lagra och söka i dokument

Retrieval Augmented Generation (RAG) med LangChain

Meri Nova

Machine Learning Engineer

Instansiera Neo4j-databasen

from langchain_community.graphs import Neo4jGraph

graph = Neo4jGraph(url="bolt://localhost:7687", username="neo4j", password="...")

import os

url = os.environ["NEO4J_URI"]
user = os.environ["NEO4J_USERNAME"]
password = os.environ["NEO4J_PASSWORD"]

graph = Neo4jGraph(url=url, username=user, password=password)

¹ https://neo4j.com/download/

Lagra grafdokument

from langchain_experimental.graph_transformers import LLMGraphTransformer

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

graph_documents = llm_transformer.convert_to_graph_documents(documents)

Lagra grafdokument

graph.add_graph_documents(
  graph_documents,

  include_source=True,

  baseEntityLabel=True

)

include_source=True: kopplar noder till källdokument via MENTIONS-kant
baseEntityLabel=True: lägger till etiketten __Entity__ på varje nod

Grafdokumenten representerade som noder och kanter.

En inzoomad version av föregående bild med noder om OpenAI-modeller.

Databasschema

print(graph.get_schema)

Node properties:
Concept {id: STRING}
Architecture {id: STRING}
Organization {id: STRING}
Event {id: STRING}
Paper {id: STRING}

The relationships:
(:Concept)-[:DEVELOPED_BY]->(:Person)
(:Architecture)-[:BASED_ON]->(:Concept)
(:Organization)-[:PROPOSED]->(:Concept)
(:Document)-[:MENTIONS]->(:Event)
(:Paper)-[:BASED_ON]->(:Concept)

Söka i Neo4j – Cypher Query Language

En nod kallad James med en relation kallad friends som pekar mot en okänd personnod.

Söka i Neo4j – Cypher Query Language

Söka i LLM-grafen

results = graph.query("""
MATCH (gpt4:Model {id: "Gpt-4"})-[:DEVELOPED_BY]->(org:Organization)
RETURN org
""")


print(results)

[{'org': {'id': 'Openai'}}]

Nu kör vi en övning!

Retrieval Augmented Generation (RAG) med LangChain