> GraphIndex: Incremental Multi-Modal KG Indexing Without LLMs
Three fused index modalities (structural, semantic, relational) with cross-modal incremental propagation: 10-19x rebuild speedups, beating BM25 on all 5 BEIR benchmarks, no LLM at index time. 169K papers at 1.16M edges.
// TABLE_OF_CONTENTS
GraphIndex: Incremental Multi-Modal KG Indexing¶
No LLM Required at Index Time¶
| System | LLM at Index | Incremental | Multi-Modal |
|---|---|---|---|
| LightRAG | Yes | Partial | No |
| Microsoft GraphRAG | Yes | No | No |
| GraphIndex | No | Yes | Yes |
LightRAG: 738s for 500 docs (with LLM). GraphIndex: 640s for 5,000 docs (no LLM). 10x throughput with 10x larger corpus.
Incremental Speedups¶
| Dataset | Speedup |
|---|---|
| SciFact (5.2K) | ~10x |
| arXiv-5K | ~14x |
| ogbn-arxiv (169K papers, 1.16M edges) | ~19x |
Speedup grows with corpus size: incremental delta is roughly constant.
Hybrid RRF Results¶
P@10 = 0.636 (BM25 + Semantic + Structural) on arXiv-5K. Semantic modality outperformed BM25 on all 5 BEIR benchmarks (p < 0.001).
RAG End-to-End (SciFact, 300 claims)¶
| Retriever | Llama-3.1-8B Accuracy |
|---|---|
| BM25 | 37.7% |
| Semantic (GraphIndex) | 43.3% |
Domain Adapter Pattern¶
244-851 LOC to add a new domain (code, scientific literature, markdown already implemented). A weekend engineering task, not a research project.
Key Lesson¶
Hybrid fusion is context-sensitive: structural modality adds value for graph-native queries but contributes little for purely semantic retrieval. Always run modality ablations.