SGK86/rag-vs-graphrag-eval

Vector-RAG vs. Graph-RAG — Multi-Hop & Multi-Needle Experiment

Controlled comparison of three retrieval pipelines over the same chunk corpus, same reader LLM, same prompts, differing only in retrieval strategy:

Pipeline	Retrieval
vector	Dense retrieval (bge-large-en-v1.5, cosine, top-5)
traversal	KG entity-match on question → 2-hop graph traversal, chunks ranked by 1/(1+hop)
ppr	Personalized PageRank with teleport to question entities (HippoRAG-style), max-score passage ranking

Setup

• Corpus: 14,067 chunks (350 words each) — MultiHop-RAG news corpus (609 docs → 3,319 chunks), 2WikiMultihopQA contexts (2,751), MuSiQue paragraphs (7,997)
• Queries: MultiHop-RAG 500 · 2Wiki 400 · MuSiQue 400 · synthetic multi-needle 90 (K∈{2,5,10} × 30, needles hidden inside full-length news chunks, combining question)
• Graph indexing: Qwen2.5-7B-Instruct extracts ≤8 triples/chunk → 14,067 LLM calls (vector pipeline needs zero LLM calls for indexing)
• Reader: Qwen2.5-7B-Instruct (greedy), identical prompt for all pipelines; context = retrieved top-5 (needle: top-max(5,K))
• Embeddings: BAAI/bge-large-en-v1.5 (truncated to 510 tokens)

Results

QA benchmarks (EM / Recall@5 / All-Recall@5 / Refusal)

Benchmark	vector	traversal	ppr
MultiHop-RAG (n=500)	31.8% / 35.9% / 9.8% / 61.8%	17.0% / 15.5% / 2.4% / 81.0%	9.6% / 9.1% / 1.0% / 88.6%
2WikiMultihopQA (n=400)	24.2% / 72.6% / 45.8% / 63.8%	14.2% / 46.7% / 20.5% / 76.2%	12.2% / 41.0% / 17.5% / 76.5%
MuSiQue (n=400)	4.8% / 45.6% / 14.5% / 82.8%	3.2% / 22.3% / 4.3% / 91.8%	3.0% / 19.4% / 2.3% / 91.0%
Aggregate (n=1300)	21.2% / 50.2% / 22.3% / 68.8%	11.9% / 27.2% / 8.5% / 82.8%	8.4% / 22.1% / 6.5% / 85.6%

Multi-needle test (K needles in distinct documents)

K	≥1 needle in context (vec/tra/ppr)	all-recall@depth	code-set-match
2	7.8% / 2.2% / 10.0%	≈0%	0%
5	—	≈0%	0%
10	—	≈0%	0%

All pipelines essentially fail to retrieve single anomalous sentences buried in full-length unrelated documents from a 14K-chunk corpus.

Diagnostics

• Graph seed matching was not the bottleneck: 100% of MultiHop/2Wiki queries matched graph entities (mean 5.9–20.5 seeds).
• EM among answered (non-refusal) is high for all pipelines (MultiHop: 83–90%) — retrieval recall, not reader correctness, is the bottleneck.
• Indexing cost: graph = 14,067 LLM extraction calls + 5.1 MB triples; vector = 14,067 embedding calls only.

Findings

1. At 7B-extractor scale, plain vector RAG dominates graph retrieval on every benchmark and every metric — while costing zero LLM indexing calls. This reproduces the independent small-model finding (GraphRAG ≈30% vs Vector-RAG ≈33% on 2Wiki with Llama-3.1-8B) and contrasts with HippoRAG's reported advantage (EM 46.6 vs 33.4, All-Recall 75.7% vs 37.1%) which relied on stronger extraction (ColBERTv2 + better OpenIE).
2. All-Recall@5 — the multi-hop bottleneck metric — is where the graph loses most: 22.3% (vector) vs 8.5%/6.5%. The KG built by a 7B model is too noisy to chain evidence reliably.
3. The needle test is failed by all corpus-level retrieval, vector included (≥1-needle-in-context only 2–10%). A single sentence inside a 350-word unrelated document barely moves the chunk embedding. Tasks that require finding distributed micro-facts are a long-context problem (in-context needles, RULER-style), not a corpus-retrieval problem.

Caveats

• KG quality is bounded by the 7B extractor (≤8 triples/chunk, no entity normalization/synonym merging); HippoRAG-grade extraction could close part of the gap.
• Entity matching is lexicon-based (question tokens ⊆ node names); graph pipelines never fell back to vector except for 1/400 MuSiQue queries, so this is not a fallback artifact.
• Single run, greedy decoding, one seed; differences of this size (>2×) are far beyond noise, but per-pipeline tuning was minimal by design.

Reproduction

Everything is in this repo: experiment.py (self-contained, stage artifacts cached in results/full/artifacts/), results/full/metrics.json + per_row.json (every prediction), retrieval.json, answers.json, state.json (chunk corpus + queries), triples.json (KG), emb.npy (embeddings).

EXP_MODE=full python experiment.py   # resumes from pushed artifacts
EXP_MODE=smoke python experiment.py  # 21-query smoke

Sources

• MultiHop-RAG: https://arxiv.org/abs/2401.15391 (dataset: yixuantt/MultiHopRAG)
• 2WikiMultihopQA: https://arxiv.org/abs/2011.01060 (dataset: framolfese/2WikiMultihopQA)
• MuSiQue: https://arxiv.org/abs/2108.00573 (dataset: dgslibisey/MuSiQue)
• HippoRAG: https://arxiv.org/abs/2405.14831