GraphRAG Reference

This page is the reference for pg_ripple's GraphRAG (Retrieval-Augmented Generation) integration.

Overview

pg_ripple provides GraphRAG capabilities that combine RDF knowledge graph retrieval with large language model (LLM) generation. The system uses pgvector-based semantic search, SPARQL query generation, RAG context assembly, and entity alignment to support knowledge-graph-grounded question answering.

Status

SELECT feature_name, status FROM pg_ripple.feature_status()
WHERE feature_name LIKE '%rag%' OR feature_name LIKE 'graphrag%';

SQL Functions

RAG Pipeline

1. Embedding: Subject IRIs are embedded via the configured LLM embedding endpoint.
2. Retrieval: Given a question, its embedding is compared to entity embeddings using HNSW/IVFFlat vector search (pg_ripple.rag_context()).
3. Context assembly: Top-k entities are expanded to their triples via SPARQL (DESCRIBE or a custom query template).
4. Generation: The assembled context is passed to the LLM for answer generation.

Microsoft GraphRAG Integration

graphrag_export() produces output compatible with Microsoft GraphRAG:

• Parquet format: entity and relationship tables with embeddings
• JSON-LD format: context-annotated graph nodes

LLM Configuration

Set LLM endpoint parameters via GUCs:

Related Pages

• Vector Search Reference
• GraphRAG Functions
• GraphRAG Ontology
• Feature Status Taxonomy