TGS-RAG Claims 80% Token Reduction vs LightRAG at Higher Accuracy

Southeast University Nanjing's TGS-RAG (Text-Graph Synergy RAG) paper claims ~80% token reduction versus LightRAG and GraphRAG on HotpotQA while achieving higher accuracy, via bidirectional verification between vector and graph retrieval channels.

1 min read|agenticonsult Intelligence

TGS-RAG Claims 80% Token Reduction vs LightRAG at Higher Accuracy

Researchers at Southeast University Nanjing have published TGS-RAG (Text-Graph Synergy RAG), a hybrid retrieval framework claiming ~217M tokens consumed on HotpotQA versus ~757M for LightRAG and ~646M for GraphRAG — a ~70–80% token reduction — while also reporting the highest accuracy in the benchmark comparison. The core innovation: beam-search graph pruning is treated as "move to memory" rather than delete, allowing the vector retrieval channel to revive discarded graph nodes when textual evidence later implicates them. The mechanism avoids new database queries during the orphan-bridging step, which is how the token savings accumulate.

Why It Matters

A 3–4× cost reduction with accuracy parity or better is a Pareto improvement — the enterprise case for adopting TGS-RAG over LightRAG/GraphRAG is one-sided if the results replicate. The infrastructure footprint (pgvector, 0.6B Qwen embedding, single 32GB Linux box) also reads as a deliberate build-it-yourself signal.

This breaking-news item was assembled from the cited primary source with AI assistance. It is intended for rapid situational awareness — refer to the original publication for the definitive statement.

TGS-RAG Claims 80% Token Reduction vs LightRAG at Higher Accuracy

TGS-RAG Claims 80% Token Reduction vs LightRAG at Higher Accuracy

Why It Matters

Live Intel Feed