Combining lexical graphs with vector search

Graph RAG with LangChain and Neo4j

Adam Cowley

Manager, Developer Education at Neo4j

Using relationships in semantic search

What is it like to work at Neo4j?

A graph diagram with a Review node and the text property "Neo4j is a great place to work"

Using relationships in semantic search

What is it like to work at Neo4j?

The review is posted by a Person with the screen name @emileifrem

Using relationships in semantic search

What is it like to work at Neo4j?

Emil Eifrem has a WORKS_AT relationship to a Company node with the name Neo4j

Text properties to embeddings

Create node embeddings from text property

a scene with HAS_CHUNK relationships to Chunk nodes with a string property

Text properties to embeddings

Create node embeddings from text property
Vector search on embeddings
Use relationships for context

The text properties on the nodes are used to create an embedding

Neo4j as a vector store

from langchain_neo4j import Neo4jVector

Neo4jVector.from_documents(): Create nodes and underlying index from Documents
Neo4jVector.from_existing_graph(): Create index on existing label and property combinations

Chunking existing node properties

A knowledge graph diagram highlighting the HAS_SCENE relationship from the Act node to the Scene node

res = graph.query("""

MATCH (s:Scene)

WHERE NOT {(s)-[:HAS_LINE]->()}

RETURN s.id AS id, s.text AS text

""")


for scene in res:
    # Create chunks from scene['text']

Chunking existing node properties

Split text by "\n\n"
Stage directions have one line
Spoken lines start with character name
- Following lines contain the spoken words

 Enter Sampson and Gregory armed with swords..


SAMPSON.

Gregory, on my word...


GREGORY.
No, for then we should...


SAMPSON.
I mean, if we be in choler, we'll draw...

Chunking existing node properties

# Split text into lines
lines = [
    line.strip() for line in text.split("\n\n")

    if "\n" in line

]


for line in lines:
    parts = line.split("\n")


    character = parts[0]

    text = "\n".join(parts[1:])

Chunking existing node properties

# ...Continuing loop over scenes

line_node = Node(type="Line", id=f"{scene.id}-line-{i}", properties={})
character_node = Node(type="Character", id=character, properties={})


# (:Scene)-[:HAS_LINE]->(:Line)
graph_document.relationships.append(
    Relationship(source=scene, target=line_node, type="HAS_LINE"))


# (:Line)-[:SPOKEN_BY]->(:Character)
graph_document.relationships.append(Relationship(
    source=line_node, target=characters[character], type="SPOKEN_BY"))

Creating a vector index on an existing graph

from langchain_openai import OpenAIEmbeddings

store = Neo4jVector.from_existing_graph(

    OpenAIEmbeddings(model="text-embedding-3-small"),

    url=NEO4J_URI,
    username=NEO4J_USERNAME,
    password=NEO4J_PASSWORD,

    node_label="Line",

    text_node_properties=["text"],

    embedding_node_property="embedding",

    index_name="lines",

)

Vector retrieval with LCEL

# Create retriever
retriever = store.as_retriever()


# Invoke manually
retriever.invoke("What does Romeo think of Juliet?") # [ Document, Document, ... ]

Let's practice!

Graph RAG with LangChain and Neo4j