← All projects

Project 04

GraphRAG Knowledge Graph

Entity extraction into Neo4j, answering multi-hop questions vectors cannot

Knowledge graphsNeo4jGraphRAGData modellingMetadata

An LLM extracts entities and relationships from the same corpus and loads them into Neo4j. Questions requiring multi-hop reasoning — which policy governs which process, who owns which upstream system — are answered by traversal rather than similarity. Runs the identical question set against both retrievers side by side, so the comparison is evidence rather than assertion.

Try it

How it works

  1. 01
    Extract

    Structured-output LLM call per chunk yields typed entities and relationships.

  2. 02
    Resolve

    Entity resolution merges duplicates by embedding similarity plus name matching.

  3. 03
    Load

    Idempotent MERGE into Neo4j AuraDB.

  4. 04
    Traverse

    Cypher generation from natural language, with a whitelist of allowed patterns.

  5. 05
    Compare

    Same questions through vector RAG and GraphRAG, scored by the project 3 harness.

Infrastructure & running cost

ServiceRoleCost
Neo4j AuraDB FreeGraph store — 200k nodes, no card required$0
LambdaExtraction + Cypher generationFree tier

Neptune Serverless would cost ~$90/mo minimum; AuraDB Free is the same skill at zero cost.