InterSystems IRIS GraphRAG

GraphRAG with InterSystems IRIS — A research and demo project that loads academic papers and their relationships into InterSystems IRIS, builds embeddings, and enables semantic + graph-based retrieval using vector search and knowledge graph queries.

Introduction

Traditional RAG pulls relevant chunks with vector search, but it can miss structure—who’s related to what, and how. GraphRAG adds that structure by blending graph context (entities + relations) with dense retrieval, so answers can be more grounded and explainable. In this demo, we use 50 recent GraphRAG-related papers (title, abstract, authors, URL, date pulled from arXiv) and a Graph Transformer to automatically extract entities (nodes) and relations (edges) from the text; the graph is stored in IRIS globals, while the same papers are embedded for dense retrieval.

This demo uses InterSystems IRIS in two complementary ways:

IRIS globals as a graph store
Globals are IRIS’s persistent, sparse, multi-dimensional arrays—perfect for graphs. We keep paper in ^GraphContent and relations in ^GraphRelations, which makes it trivial to fetch nodes/edges for any subset of documents.

IRIS SQL for vector retrieval
Each paper is embedded with MiniLM-L6-v2 (384d), stored in IRIS as VECTOR, and indexed with HNSW for fast approximate nearest-neighbor search.

Demo Overview

This app lets you run the same question through three pipelines and compare results side-by-side—LLM (no retrieval), RAG (vector retrieval over papers), and GraphRAG (RAG + a graph neighborhood from IRIS globals) with an interactive entity-relation visualization.

Here’s a simple query flow starting with the agent and showing the three paths:

User Question
      │
      ▼
 ┌───────────────┐
 │  Agent Router │  (classifies intent & picks mode / params)
 └──────┬────────┘
        │
   ┌────┼───────────────┬─────────────────────┐
   │    │               │                     │
   │    ▼               ▼                     ▼
   │  LLM            RAG                GraphRAG
   │ (no retrieval)  (vector-only)      (vector + graph)
   │                 ┌───────────────┐  ┌──────────────────┐
   │                 │ Embed query   │  │ Embed query      │
   │                 │ MiniLM-L6-v2  │  │ MiniLM-L6-v2     │
   │                 └───────┬───────┘  └─────────┬────────┘
   │                         ▼                    ▼
   │                 ANN search (HNSW)      ANN search (HNSW)
   │                 over IRIS SQL          over IRIS SQL
   │                 table `paper_content`  table `paper_content`
   │                         │                    │
   │                         ▼                    ▼
   │                  Top-K papers           Doc IDs → fetch
   │                                          nodes/edges from
   │                                          IRIS globals:
   │                                          ^GraphContent,
   │                                          ^GraphRelations
   │                         │                    │
   │                         ▼                    ▼
   │                  Generate answer        Build subgraph →
   │                                         Generate answer
   │                                         + graph visualization
   └──────────────────────┼──────────────────────────┼────────
                          ▼                          ▼
                     UI shows answer            UI shows answer
                                                + entity-relation
                                                visualization

InterSystems IRIS GraphRAG Quickstart

Clone the repo

      git clone https://github.com/fanji-isc/IRIS-Global-GraphRAG.git

Add your own OPEN AI key to .env file in app/
```
      OPENAI_API_KEY=sk-...
```
Start the Docker containers (one for IRIS, one for Flask):
```
     docker compose up --build
```
Open the app & IRIS

Demo: http://localhost:8000

IRIS Portal: http://localhost:52773

InterSystems IRIS GraphRAG Works