End-to-End GraphRAG Architecture

Memgraph holds the knowledge graph alongside vector and text indexes in one engine, so retrieval can combine them. An agent reads the question, picks a retrieval strategy, and Memgraph runs it as a single query. Four steps:

1. 1

   Build the knowledge graph

   Documents and records become entities and relationships. Every node keeps a reference back to its source.

   PDFs, SQL, apps, KBs

   Modelling + entity resolution

   Unstructured2Graph, SQL2Graph

   Memgraph

   graph, vector and text indexes, source refs

2. 2

   Route the question and pick a strategy

   Agentic GraphRAG

   The agent classifies the question and picks a strategy. Each one is a single Cypher query that Memgraph executes. Combine freely, or add your own.

   Semantic retrieval

   vector or text search

   “What does the leave policy say?” Vector search for meaning, text search for exact terms and identifiers. No traversal needed.

   Structured analysis

   Text2Cypher

   “How many policies have open exceptions?” A validated, read-only Text2Cypher query counts, filters and traverses relationships without semantic search.

   Local graph search

   search + traversal

   “Which rules apply to this employee?” Search finds the entry point, the graph expands to what's connected, then ranking trims it.

   Global approach

   query focused summarization

   “What changed across last year's revisions?” Retrieve relevant precomputed community summaries and synthesize an answer across them.

3. 3

   Return context the answer can cite

   The query returns a bounded result rather than a pile of chunks, and each fact still carries where it came from.

   Retrieved context + sources

   document, page, system, record

   LLM

   reasons over what it was given

4. 4

   Answer, with the evidence attached

   Return source identifiers with the retrieved context, require the model to cite them, and validate that cited evidence supports the generated claims.