AI Insights · Oct 22, 2026 · Verified
Search first, graphs only for multi-step questions
Combined keyword and vector search, with an agent that searches again when needed, is the baseline. Knowledge graphs only pay off on questions that need several linked facts.
Claim
Three neutral studies converge: plain RAG wins single-hop questions, graph methods win multi-hop and query-focused summarisation, and an agentic search loop over dense retrieval closes most of the gap. A lightweight graph (personalised PageRank over an entity and passage graph) is the only graph variant that beats strong embedders at reasonable cost. Community-summary GraphRAG and LightRAG do not.
Evidence
- HippoRAG 2 (ICML 2025), Table 2 with Llama-3.3-70B reader, F1: HippoRAG 2 59.8 average versus NV-Embed-v2 57.0, GraphRAG 49.6, RAPTOR 48.8, LightRAG 6.6. Gains concentrate on multi-hop (52.0 versus 48.1 for the best embedder); simple QA is a wash. LightRAG was run at defaults, and its own EMNLP 2025 Findings paper reports favourable results, so that comparison is contested.
- Han et al. (v3 2026-03-04): unified protocol over QA and summarisation; RAG and GraphRAG have "distinct strengths" and a selection or hybrid strategy gives consistent gains. No dollar costs published.
- Fan et al., "Do We Still Need GraphRAG?" (2026-04-01): inside agentic search loops, prompted and RL-trained, agentic search "substantially improves dense RAG and narrows the performance gap to GraphRAG". GraphRAG keeps an advantage on multi-hop and more stable agent behaviour once its offline cost is amortised.
- Graphiti (32.1k stars) is the most-adopted incremental graph for agents: bi-temporal edges with automatic fact invalidation, hybrid semantic plus BM25 plus traversal retrieval, episode provenance, Neo4j 5.26+ or FalkorDB backends. Its MCP server has no auth and is not production-grade. Zep's benchmark numbers are self-reported.
Open questions
- No study covers scientific-paper corpora specifically. The multi-hop advantage may be larger there because citation chains are inherently multi-hop.
Sources
- https://arxiv.org/abs/2502.14802 — paper, Wed Jun 18
- https://arxiv.org/abs/2502.11371 — paper, Tue Mar 03
- https://arxiv.org/abs/2604.09666 — paper, Tue Mar 31
- https://aclanthology.org/2025.findings-emnlp.568/ — paper, Fri Oct 31
- https://github.com/getzep/graphiti — repo, Wed Oct 21
AI-assisted research: gathered and summarised with AI from the sources above, each claim checked against them.