AI/Embedding/Graph/Concept/How different from RAG

• Graph Embeddings in Neo4j — Conceptual Overview

  • Overview

    • Graph embeddings are vectors that encode the structure and attributes of graph elements so that geometric proximity mirrors topological similarity.

  • Context

    • Neo4j’s Graph Data Science (GDS) library learns these vectors and stores them on nodes, enabling fast similarity search, link prediction, and hybrid retrieval-augmented generation (RAG) that combines structured and unstructured knowledge.

  • Key Principles

    • Topology-aware representation — embeddings capture paths, neighborhoods, and community structure rather than raw text.
    • Algorithmic generation — vectors come from graph algorithms such as FastRP, Node2Vec, and GraphSAGE, not from large language models.
    • Native storage & indexing — Neo4j writes embeddings back as LIST<FLOAT> properties and can build vector indexes for k-NN queries.
    • Complement to text embeddings — graph embeddings support relationship reasoning, while text embeddings capture semantic content; combining both powers “GraphRAG.”

  • Mechanism

    • FastRP projects higher-order adjacency information into a low-dimensional space using random projections, delivering production-scale speed.
    • Node2Vec performs biased random walks and trains a skip-gram model so nodes that co-occur in walks land near each other.
    • GraphSAGE samples neighbors and aggregates their features through a neural network, allowing inductive embeddings for unseen nodes.
    • After computation, embeddings can be indexed:
      • CALL gds.fastRP.mutate('g', {embeddingDimension:128, mutateProperty:'embedding'});
      • CREATE VECTOR INDEX ON :Person(embedding);

  • Examples

    • cypher

      // 1. Project a social graph
      CALL gds.graph.project(
          'social',
          'Person',
          {FRIEND_OF: {orientation:'UNDIRECTED'}}
      );
      // 2. Generate FastRP embeddings
      CALL gds.fastRP.mutate(
          'social',
          {embeddingDimension:128, mutateProperty:'embedding'}
      );
      // 3. Find nearest neighbours
      CALL gds.knn.stream(
          'social',
          {topK:5, nodeProperties:['embedding']}
      ) YIELD node1, node2, similarity;

  • Misconceptions

    • “Graph embeddings are just text embeddings stored in a graph.”

      – They are learned from topology and attributes, independent of language models.
      

    • “Embeddings remove the need for Cypher pattern matching.”

      – They approximate similarity but cannot replace exact path queries where relationship semantics matter.
      

  • Related

    • AI/RAG/Graph/Graph Vector

    • Neo4j/Blog/24/09/Enhancing Hybrid Retrieval with Graph Traversal Using the GraphRAG Python Package
    • Fast Random Projection – Neo4j Graph Data Science (Graph Database & Analytics)
    • Node2Vec – Neo4j Graph Data Science (Graph Database & Analytics)
    • GraphSAGE – Neo4j Graph Data Science (Graph Database & Analytics)
    • Vector indexes – Cypher Manual (Graph Database & Analytics)