CodeAlchemy

Jotting one man's journey through software development, programming, and technology

Project maintained by pablogarciaprado Hosted on GitHub Pages — Theme by mattgraham

Knowledge Graphs

◀️ Home

What is a Knowledge Graph?

A knowledge graph is a database that stores information in nodes and relationships.

Both nodes and relationships can have properties. Nodes can be given labels, which are a way of grouping multiple nodes together. Relationships always have a type and a direction.

How to Build a Knowledge Graph

1. Start with your use case

If you don’t anchor the graph in real questions, it’ll turn into an expensive diagram.

Examples:

“Which customers are at risk of churn?”
“Who reports to whom across departments?”
“What products are related and why?”

This step comes from principles in Knowledge Representation. You’re modeling reality for a purpose, not for completeness.

2. Identify entities and relationships

Entities (nodes): these are the “things”.

Person
Company
Product
Document

Relationships (edges): connect entities:

works_at
manages
purchased
depends_on

Example:

Pablo ──works_at──> CompanyX
Pablo ──manages──> Olbap

Keep it simple at first. Start with a Minimum Viable Graph (MVG), then extract, enhance, expand, and repeat to grow the graph. Over-modeling is a common mistake.

3. Define a schema (ontology)

This is your structure:

What types of entities exist?
What relationships are allowed?
What properties each entity has?

4. Assign unique identifiers

Every entity should have a unique ID.

5. Ingest your data

This is where most of the real work happens.

Structured data: Databases (CRM, ERP), CSV files
Unstructured data: PDFs, emails, docs

6. Store it in a graph database

You need a system designed for relationships.

Popular options that let you query relationships efficiently:

Neo4j
Amazon Neptune
ArangoDB

Neo4j is a graph database management system designed to store, query, and analyze data as networks of nodes and relationships rather than in traditional tables.

7. Query the graph

Instead of SQL, you’ll use graph query languages like Cypher (Neo4j).

Example query:

“Find all employees managed by Pablo who work in Sales”

8. Connect it to your AI assistant

Use the graph to answer structured queries directly and/or combine it with RAG.

RAG vs Knowledge Graphs

RAG (Retrieval-Augmented Generation)

Built on vector search, typically with embeddings from Natural Language Processing. It’s best at:

Searching unstructured data (docs, PDFs, emails, wikis)
Answering natural-language questions
Fast implementation and iteration

Strengths

Simple architecture
Works well with messy/internal documents
Flexible—handles vague queries

Weak spots

Can miss relationships across documents
Less reliable for multi-step reasoning
Harder to enforce strict correctness

Knowledge Graph

Structured data with entities and relationships (nodes + edges), often used in Knowledge Representation. It’s best at:

Complex relationships (“Who reports to whom across departments?”)
Logical queries and reasoning
Situations where correctness and traceability matter

Strengths

Precise, explainable answers
Great for multi-hop reasoning
Strong data governance

Weak spots

Expensive to build and maintain
Requires structured/clean data
Slower to get started

A practical decision framework

1. What does your data look like?

Mostly documents, PDFs, Slack, Notion → RAG
Structured entities (customers, products, org charts) → Knowledge graph

2. What kind of questions will users ask?

“What’s our refund policy?” → RAG
“Which customers bought product X and later upgraded within 6 months?” → Knowledge graph

3. Do relationships matter deeply?

If your assistant needs to understand chains like:

employee → manager → department → budget

then knowledge graphs shine.

If it’s mostly:

“find relevant info and explain it”

RAG is enough.

4. How important is correctness vs speed?

Need fast MVP → RAG
Need auditable, deterministic answers → Knowledge graph

Hybrid approach

Often the sweet spot is not having to pick one:

Use RAG for:

Document retrieval
General Q&A

Use knowledge graph for:

Core business entities
Relationship-heavy queries

Let the assistant decide which to query.

Graph → precise relationships RAG → unstructured knowledge

Concrete examples

RAG:

Internal company chatbot
Support assistant
Documentation search tool

Knowledge graphs:

Fraud detection system
Enterprise analytics assistant
Compliance or legal reasoning system