Chunk size

The graph created by the SimpleKGPipeline is based on chunks of text extracted from the documents. By default, the chunk size is quite large, which may result in fewer, larger chunks. The larger the chunk size, the more context the LLM has when extracting entities and relationships, but it may also lead to less granular data.

In this lesson, you will modify the SimpleKGPipeline to use a different chunk size.

Delete the existing graph

You will be re-importing the data and modifying the existing graph. To ensure a clean state, you can delete the graph at any time using:

cypher

Delete the existing graph

MATCH (n) DETACH DELETE n

Text Splitter Chunk Size

To modify the chunk size you will need to create a FixedSizeSplitter object and pass it to the SimpleKGPipeline when creating the pipeline instance:

Modify the genai-graphrag-python/kg_builder.py file to import the FixedSizeSplitter class and create an instance with a chunk size of 500 characters:
python
```
from neo4j_graphrag.experimental.components.text_splitters.fixed_size_splitter import FixedSizeSplitter

text_splitter = FixedSizeSplitter(chunk_size=500, chunk_overlap=100)
```
The chunk_size parameter defines the maximum number of characters in each text chunk. The chunk_overlap parameter ensures that there is some overlap between consecutive chunks, which can help maintain context.

Update the SimpleKGPipeline instantiation to use the custom text splitter:

python

kg_builder = SimpleKGPipeline(
    llm=llm,
    driver=neo4j_driver, 
    neo4j_database=os.getenv("NEO4J_DATABASE"), 
    embedder=embedder, 
    from_pdf=True,
    text_splitter=text_splitter,
)

Reveal the complete code

python

import os
from dotenv import load_dotenv
load_dotenv()

import asyncio

from neo4j import GraphDatabase
from neo4j_graphrag.llm import OpenAILLM
from neo4j_graphrag.embeddings import OpenAIEmbeddings
from neo4j_graphrag.experimental.pipeline.kg_builder import SimpleKGPipeline
from neo4j_graphrag.experimental.components.text_splitters.fixed_size_splitter import FixedSizeSplitter

neo4j_driver = GraphDatabase.driver(
    os.getenv("NEO4J_URI"),
    auth=(os.getenv("NEO4J_USERNAME"), os.getenv("NEO4J_PASSWORD"))
)
neo4j_driver.verify_connectivity()

llm = OpenAILLM(
    model_name="gpt-4o",
    model_params={
        "temperature": 0,
        "response_format": {"type": "json_object"},
    }
)

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

text_splitter = FixedSizeSplitter(chunk_size=500, chunk_overlap=100)

kg_builder = SimpleKGPipeline(
    llm=llm,
    driver=neo4j_driver, 
    neo4j_database=os.getenv("NEO4J_DATABASE"), 
    embedder=embedder, 
    from_pdf=True,
    text_splitter=text_splitter,
)

pdf_file = "./genai-graphrag-python/data/genai-fundamentals_1-generative-ai_1-what-is-genai.pdf"
result = asyncio.run(kg_builder.run_async(file_path=pdf_file))
print(result.result)

Run the modified pipeline to recreate the knowledge graph with the new chunk size.

cypher

View the documents and chunks

MATCH (d:Document)<-[:FROM_DOCUMENT]-(c:Chunk)
RETURN d.path, c.index, c.text
ORDER BY d.path, c.index

You can experiment with different chunk sizes to see how it affects the entities extracted and the structure of the knowledge graph.

cypher

View the entities extracted from each chunk

MATCH p = (c:Chunk)-[*..3]-(e:__Entity__)
RETURN p

Check your understanding

What is the primary trade-off when increasing the chunk size in the SimpleKGPipeline?

❏ Larger chunks process faster but use more memory
✓ Larger chunks provide more context for entity extraction but result in less granular data
❏ Larger chunks create more entities but fewer relationships
❏ Larger chunks improve accuracy but require more computational power

Hint

Consider what happens to the level of detail and context when you make text chunks bigger or smaller.

Solution

The larger the chunk size, the more context the LLM has when extracting entities and relationships, but it may also lead to less granular data. This is the key trade-off - more context versus granularity of the extracted information.

Lesson Summary

In this lesson, you:

Learned about the impact of chunk size on entity extraction
Modified the SimpleKGPipeline to use a custom chunk size with the FixedSizeSplitter

In the next lesson, you will define a custom schema for the knowledge graph.

Constructing Knowledge Graphs with Neo4j GraphRAG for Python

Introduction

Knowledge Graph Pipeline

Retrieval

Customization

Chunk size

Delete the existing graph

Text Splitter Chunk Size

Check your understanding

What is the primary trade-off when increasing the chunk size in the SimpleKGPipeline?

Lesson Summary

Chatbot