Knowledge Graph — What It Is

What It Does

The Knowledge Graph uses LLMs (OpenAI) to extract entities, decisions, relationships, and constraints from natural-language documents. Results are persisted to Neo4j for rich graph queries.

Example

Given a document about authentication architecture, the KG extracts:

Entity: AuthService (component)
Entity: JWT (technology)
Entity: Session-based auth (option)
Relationship: AuthService USES JWT
Relationship: Team DECIDED_FOR JWT
Relationship: Team DECIDED_AGAINST Session-based auth

You can then query:

iw query "What decisions were made about authentication?"

# → JWT was chosen over session-based auth for the authentication system
# → Reasons: statelessness, horizontal scaling, API compatibility

How It Works

Pipeline: IN → FX → KX → GX

IN — Chunks documents (semantic markdown splitting, ~16k chars/chunk)
FX — LLM extracts raw entity-relationship triples per chunk
KX — Canonicalizes entities and predicates into a consistent schema (30 predicate types)
GX — Cross-document entity deduplication (exact + fuzzy merge)

What Gets Extracted

Entity types: concept, decision, option, requirement, feature, component, technology, resource, role, risk, phase, constraint, question, tradeoff

Relationship types: CONTAINS, DEPENDS_ON, IMPLEMENTS, EXTENDS, CALLS, USES, DECIDED_FOR, DECIDED_AGAINST, SUPERSEDES, ENABLES, BLOCKS, RISKS, and more.

When To Use It

Need	Use
Find relevant files	CARI (free, fast)
Check doc freshness	CARI (free, fast)
Understand decisions and rationale	KG
Trace impact of architectural changes	KG
Query natural-language questions	KG
Build RAG context for AI agents	KG

Requirements

Dependency	Purpose
Neo4j 5	Graph storage (Docker recommended)
OpenAI API key	LLM extraction + NL queries

Next Steps

Try It — set up Neo4j and run your first extraction
CLI Reference — full command documentation