BridgeRAG: Training-Free Bridge-Conditioned Retrieval
for Multi-Hop Question Answering

Andre Bacellar

Preprint, 2026

BridgeRAG ranks later-hop evidence by bridge-conditioned utility rather than query-only relevance — without a graph database, proposition index, or any training.
Paper PDF arXiv GitHub Artifacts BibTeX

Results

Best published training-free R@5 under matched benchmark evaluation on all three standard MHQA benchmarks.

Method MuSiQue R@5 2WikiMultiHop R@5 HotpotQA R@5 Training Graph
HippoRAG2 0.747 0.904 0.963 No Yes (PPR)
PropRAG 0.783 0.941 0.974 No Yes (prop.)
BridgeRAG (ours) 0.8146 (+3.1pp) 0.9527 (+1.2pp) 0.9875 (+1.35pp) No No

Δ vs. PropRAG. R@5 = mean(|gold ∩ top-5| / |gold|). All methods use NV-Embed-v2 embeddings.

The Core Idea

Standard retrieval scores every candidate by s(q, c) — similarity to the query alone. For multi-hop questions this is the wrong target. The second-hop passage is relevant not to the original query but to what the first hop already found.

BridgeRAG replaces s(q, c) with s(q, b, c), where b is the top-1 first-hop "bridge" passage. A tripartite LLM judge attends jointly to the query and the bridge to identify which reasoning chain is active before ranking second-hop candidates.

Motivating Example

Query

"Who is the spouse of the actor who played the Terminator?"

2-way judge s(q, c) — wrong

"The Terminator is a 1984 film directed by James Cameron…" — film passage, high surface overlap, wrong chain.

Bridge b — top hop-1 passage

"Arnold Schwarzenegger is an Austrian-American actor and former politician…"

3-way judge s(q, b, c) — correct ✓

"Maria Shriver is an American journalist and author, and the wife of Arnold Schwarzenegger…"

What Is New

Mechanism Validation

Four controlled experiments — all reusing cached artifacts (zero extra LLM calls for three of four) — show the conditioning signal is:

Selective +2.55pp on parallel-chain queries (bridge_comparison, p<0.001, Bonferroni-corrected), ≈0 on single-chain subtypes. The effect appears exactly where chain ambiguity is highest.
Irreplaceable Substituting the retrieved bridge with SVO query text reduces R@5 by 2.1pp on MuSiQue — performing worse than the hardest-to-retrieve pool passage. Passage content, not retrieval intent, drives the gain.
Predictable cos(bridge, gold_hop2) positively correlates with per-query B→C improvement (Spearman ρ=0.104, p<0.001 on MuSiQue). Queries where the bridge is closer to the gold passage benefit more from conditioning.
Mechanistically precise Bridge conditioning causes productive re-rankings: 18.7% of judge-induced top-1 changes improve R@5 on parallel-chain queries vs. 0.6% on single-chain queries. Not noise — correct decisions.

Reproducibility

Models: NV-Embed-v2 (embeddings) · Llama 3.3 70B AWQ (LLM judge, SVO generation, entity extraction)

Infrastructure: vLLM on NVIDIA A100 80 GB · pgvector ANN index · fully local, no closed-source API calls

Cost per query: 3 LLM calls · 6 ANN passes · ~4–5 seconds sequential on our evaluation server

Artifacts: Canonical results at eval/pilot_audit/bridgerag_canonical_results.json. Provenance for the MuSiQue result (three independent runs, 0.8138–0.8146, LLM non-determinism) documented in ARTIFACTS.md.

BibTeX

@article{bacellar2026bridgerag,
  title     = {{BridgeRAG}: Training-Free Bridge-Conditioned Retrieval
               for Multi-Hop Question Answering},
  author    = {Bacellar, Andre},
  journal   = {arXiv preprint arXiv:2604.03384},
  year      = {2026}
}

Andre Bacellar · 2026 · arXiv · GitHub