GitHub Repository

Graph RAG with pure vector search, achieving SOTA performance in multi-hop reasoning scenarios.

202 starsPython

Replaced Neo4j with pure vector search for Graph RAG

Name: Replaced Neo4j with pure vector search for Graph RAG
Availability: InStock
Author: zhangchen

by zhangchen·Apr 8, 2026·2 points·0 comments

Visit Project View on HN

AI Analysis

●●SolidBig BrainDark Horse

Graph RAG without Neo4j — pure vector search beats HippoRAG on multi-hop benchmarks.

Strengths

•Encodes entities and relations as vectors — no separate graph database infrastructure needed.
•Single-pass reranking replaces iterative LLM agent loops, cutting computational cost.
•87.8% Recall@5 on multi-hop QA benchmarks with Milvus Lite zero-config setup.

Weaknesses

•Graph RAG is a crowded category with Microsoft's implementation and many alternatives.
•Tied to Milvus ecosystem — less flexible than database-agnostic approaches.

Similar Projects

AI/ML●●Solid

PardusDB – SQLite-like vector database in Rust

Using a single-file .pardus format with CREATE/INSERT/SELECT + SIMILARITY queries gives a very familiar developer UX for embedding storage. The combination of graph-based ANN, full transactions, thread-safety, and zero external dependencies is an uncommon and useful engineering combo for local-first AI work; it would win more attention with benchmark comparisons and richer ecosystem integrations (connectors/clients).

Niche GemWizardry

JasonHEIN

204mo ago

Infrastructure●●Solid