Building a Graph

The Problem with Traditional RAG

Traditional RAG systems work - but they’re blind to context:

The Challenge with PDF Documents: - Rich information about companies, financials, and risks locked in unstructured text - Hard to search, query, or analyze systematically - Connections between entities are hidden in narrative text - Traditional RAG can’t reason across relationships

The GraphRAG Solution: - Use AI to extract structured entities and relationships - Create a knowledge graph that preserves connections - Give the system a "mental map" of your domain - Enable context-aware retrieval, not just similarity search

It’s like giving someone index cards with code snippets vs. an architectural diagram - GraphRAG understands the structure.

From PDF Documents to Knowledge Graph

// See what documents were processed and how many chunks each has
MATCH (d:Document)<-[:FROM_DOCUMENT]-(c:Chunk)
RETURN d.path, count(c) as totalChunks
ORDER BY totalChunks DESC

Step 3: The GraphRAG Pipeline

The complete pipeline orchestrated the transformation from PDF to knowledge graph using AI-powered extraction.

The GraphRAG Pipeline:

pipeline = SimpleKGPipeline(
    driver=driver, # Neo4j connection driver
    llm=llm, embedder=embedder,  # OpenAI llm and embeddings
    entities=entities, relations=relations,  # Define schema
    enforce_schema="STRICT",
    prompt_template=prompt_template,
)
# Process the SEC filing documents
pdf_documents = [
    "apple-10K-2023.pdf", "microsoft-10K-2023.pdf",
    # ... more company filings
]
# Run the pipeline to transform PDFs into knowledge graph
for pdf_file in pdf_documents:
    pipeline.run(file_path=pdf_file)

// Count what entities were extracted by type
MATCH (e)
WHERE NOT e:Document AND NOT e:Chunk
RETURN labels(e) as entityType, count(e) as count
ORDER BY count DESC

Step 4: Adding Structured Data

But PDF extraction was only part of the story. The knowledge graph also includes structured data loaded from CSV files to complement the extracted PDF entities.

Structured Data Sources:

Why Both Data Types?

This created a complete picture: detailed company information from PDFs plus structured ownership and filing relationships. The bridge between structured and unstructured data enables the powerful GraphRAG queries you’ll explore.

// See the complete data model - all node types
MATCH (n)
RETURN labels(n) as nodeType, count(n) as count
ORDER BY count DESC

CALL db.schema.visualization()

// See how all three data types connect for one company
MATCH (c:Company {name: 'APPLE INC'})
OPTIONAL MATCH (c)-[r1]->(extracted)
WHERE NOT extracted:Chunk AND NOT extracted:Document
OPTIONAL MATCH (am:AssetManager)-[r2:OWNS]->(c)
OPTIONAL MATCH (c)<-[:MENTIONS]->(chunk:Chunk)
RETURN c.name,
       count(DISTINCT extracted) as extractedEntities,
       count(DISTINCT am) as assetManagers,
       count(DISTINCT chunk) as textChunks

// Count what the pipeline created
MATCH (d:Document)
OPTIONAL MATCH (d)<-[:FROM_DOCUMENT]->(c:Chunk)
OPTIONAL MATCH (c)<-[:FROM_CHUNK]-(e)
RETURN d.path,
       count(DISTINCT c) as chunks,
       count(DISTINCT e) as entities
ORDER BY entities DESC

// See all asset managers that were loaded
MATCH (am:AssetManager)
RETURN am.managerName, count{(am)-[:OWNS]->()} as companiesOwned
ORDER BY companiesOwned DESC
LIMIT 10

// Check data quality across companies
MATCH (c:Company)
OPTIONAL MATCH (c)-[r]->(entity)
RETURN c.name, count(r) as totalRelationships,
       collect(DISTINCT type(r)) as relationshipTypes
ORDER BY totalRelationships DESC
LIMIT 5

// Find all financial metrics for a specific company
MATCH (c:Company {name: 'MICROSOFT CORP'})-[:HAS_METRIC]->(m:FinancialMetric)
RETURN c.name, m.name
LIMIT 10

// Discover risk factors across all companies
MATCH (c:Company)-[:FACES_RISK]->(r:RiskFactor)
RETURN c.name, r.name
LIMIT 50

Key Takeaways

✅ Unstructured → Structured: PDF text was transformed into business entities and relationships

✅ Schema-Driven: Clear entity definitions guided accurate extraction

✅ AI-Powered: LLMs identified and extracted meaningful business concepts

✅ Relationship-Aware: Connections between entities were preserved and made explicit

✅ Data Model Ready: Clean, structured data prepared for the knowledge graph you’ll explore

This structured data model is the foundation for everything that follows - without it, you’d still have unstructured text instead of the queryable business entities you’ll work with!

Continue