Vector + Cypher Retriever

Create the Vector Index

You will need to create a vector index on the Chunk nodes embedding properties:

CREATE VECTOR INDEX chunkEmbedding IF NOT EXISTS
FOR (n:Chunk)
ON n.embedding
OPTIONS {indexConfig: {
 `vector.dimensions`: 1536,
 `vector.similarity_function`: 'cosine'
}};

A vector index named chunkEmbedding will be created for nodes with a Chunk label, indexing the embedding property. The index is configured to 1536 dimensions (as provided by the text-embedding-ada-002 embedding model) and use cosine similarity for distance calculations.

You can search the vector index by creating an embedding for a search term:

WITH genai.vector.encode(
    "Retrieval Augmented Generation",
    "OpenAI",
    { token: "sk-..." }) AS userEmbedding
CALL db.index.vector.queryNodes('chunkEmbedding', 5, userEmbedding)
YIELD node, score
RETURN node.text, score

Create a Vector + Cypher GraphRAG pipeline

The neo4j_graphrag package includes a VectorCypherRetriever class that combines vector similarity search with Cypher retrieval.

You can use this retriever to create a GraphRAG pipeline to:

Open genai-graphrag-python/vector_cypher_rag.py and review the code:

import os
from dotenv import load_dotenv
load_dotenv()

from neo4j import GraphDatabase
from neo4j_graphrag.embeddings.openai import OpenAIEmbeddings
from neo4j_graphrag.retrievers import VectorCypherRetriever
from neo4j_graphrag.llm import OpenAILLM
from neo4j_graphrag.generation import GraphRAG

# Connect to Neo4j database
driver = GraphDatabase.driver(
    os.getenv("NEO4J_URI"), 
    auth=(
        os.getenv("NEO4J_USERNAME"), 
        os.getenv("NEO4J_PASSWORD")
    )
)

# Create embedder
embedder = OpenAIEmbeddings(model="text-embedding-ada-002")

# Define retrieval query
retrieval_query = """
RETURN node.text as text, score
"""

# Create retriever
retriever = VectorCypherRetriever(
    driver,
    neo4j_database=os.getenv("NEO4J_DATABASE"),
    index_name="chunkEmbedding",
    embedder=embedder,
    retrieval_query=retrieval_query,
)

#  Create the LLM
llm = OpenAILLM(model_name="gpt-4o")

# Create GraphRAG pipeline
rag = GraphRAG(retriever=retriever, llm=llm)

# Search
query_text = "Where can I learn more about knowledge graphs?"

response = rag.search(
    query_text=query_text, 
    retriever_config={"top_k": 5},
    return_context=True
)

print(response.answer)

# Close the database connection
driver.close()

The retriever is configured to use the chunkEmbedding vector index you just created.

retriever = VectorCypherRetriever(
    driver,
    neo4j_database=os.getenv("NEO4J_DATABASE"),
    index_name="chunkEmbedding",
    embedder=embedder,
    retrieval_query=retrieval_query,
)

When you run the code:

You can print the context passed to the LLM by adding the following to the end of the code:

print("CONTEXT:", response.retriever_result.items)

Retrieval Cypher Query

The VectorCypherRetriever also allows you to define a Cypher query to retrieve additional context from the knowledge graph.

Adding additional context can help the LLM generate more accurate responses.

Update the retrieval_query to add additional information about the lessons, technologies, and concepts related to the chunks:

retrieval_query = """
RETURN DISTINCT
    node.text as text, score,
    collect { MATCH (node)-[:FROM_DOCUMENT]->(d)-[:PDF_OF]->(lesson) RETURN lesson.url} as lesson_url,
    collect { MATCH (node)<-[:FROM_CHUNK]-(e:Technology) RETURN e.name } as technologies,
    collect { MATCH (node)<-[:FROM_CHUNK]-(e:Concept) RETURN e.name } as concepts
"""

The retriever will execute the Cypher query adding more context.

Running the code again for the same query, "Where can I learn more about knowledge graphs?", will produce a more detailed response:

You can learn more about knowledge graphs in the Neo4j blog post linked here: What Is a Knowledge Graph?. Additionally, you can explore further lessons on knowledge graphs on the GraphAcademy website, specifically in the course "GenAI Fundamentals," including the sections "What is a Knowledge Graph" and "Creating Knowledge Graphs."

The retrieval query includes additional context relating to technologies and concepts mentioned in the chunks.

Experiment asking different questions relating to the knowledge graph such as "What technologies and concepts support knowledge graphs?".