// the find
HKUDS/MiniRAG
[ACL2026] "MiniRAG: Making RAG Simpler with Small and Open-Sourced Language Models"
MiniRAG is a graph-based RAG framework from HKUDS (the same group behind LightRAG) that builds a heterogeneous graph combining text chunks and named entities to help small language models (1.5B–4B params) do retrieval without needing strong semantic understanding. It's aimed at on-device or resource-constrained scenarios where you can't run GPT-4 class models. The ACL 2026 acceptance signals it went through peer review, not just a preprint dump.
The core idea is sound: offload semantic reasoning into graph topology so the LLM only has to answer, not navigate. The benchmark numbers show real gains over NaiveRAG and LightRAG when using SLMs — 10+ percentage point improvements on Phi-3.5-mini are not noise. The storage claim (25% of LLM-based methods) is specific and testable. Backend support is surprisingly wide — Neo4j, PostgreSQL, Redis, Milvus, Chroma, Weaviate, MongoDB all have implementations, so you're not locked in.
The package literally tells you to install `lightrag-hku` to get MiniRAG, which means this is closer to a thin wrapper and research artifact than a standalone library — the abstraction boundary isn't clean. The LiHua-World benchmark is synthetic chat logs from a fictional user, which is a controlled scenario that may not reflect real document corpora. The repo has `__pycache__` committed and `.cpython-39.pyc` files checked in, which is a signal the codebase hygiene is research-grade, not production-grade. Last push was October 2025 and activity looks like it's winding down post-publication.