raphaelmansuy/edgequake

What it solves

EdgeQuake is a high-performance Graph-RAG framework designed to overcome the limitations of traditional vector-based RAG. While standard RAG relies on similarity searches that often miss structural relationships, EdgeQuake transforms documents into intelligent knowledge graphs, enabling superior multi-hop reasoning and the ability to answer thematic or relationship-based questions.

How it works

Built in Rust for high performance, the system implements the LightRAG algorithm. It decomposes documents into a knowledge graph consisting of entities and relationships. During retrieval, it traverses both the vector space (using pgvector) and the graph structure (using Apache AGE), combining embedding-based similarity with graph traversal to provide more comprehensive answers.

Who it’s for

Developers and organizations needing a production-ready RAG system that can handle complex queries across large document sets, requiring high concurrency, low latency, and robust multi-tenant isolation.

Highlights

• Six Query Modes: Offers Naive, Local, Global, Hybrid, Mix, and Bypass modes to balance latency and depth of retrieval.
• PDF Vision Pipeline: Supports both fast text extraction and a vision-based mode (using GPT-4o, Claude, or Gemini) to reconstruct complex tables and multi-column layouts.
• Production-Grade Infrastructure: Features a REST API with SSE streaming, built-in authentication, and support for PostgreSQL 16/17/18.
• High Efficiency: Claims significant improvements over traditional RAG in query latency, entity extraction volume, and memory usage per document.
• Agent Integration: Exposes capabilities to AI agents via the Model Context Protocol (MCP).