Aethel: Graph-Retrieval Framework Achieves 100% Recall on Multi-Hop QA

Aethel addresses the challenge of synthesizing fragmented financial disclosures for secondary private equity transactions, where key metrics are scattered across documents with little lexical overlap. The framework models corpora as entity-passage graphs, propagating relevance via bipartite Personalized PageRank. A novel Bipartite Coreference Teleportation layer resolves coreferences to strengthen entity links, while a specialist-agent architecture orchestrates retrieval and reasoning. Evaluated on 200-question samples from 2WikiMultiHopQA and MuSiQue, Aethel achieves HR@5 of 100.0% and 88.5% respectively, improving coverage over vanilla graph retrieval but trading off top-rank precision.

On a real-world 4,123-chunk financial disclosure corpus, graph retrieval outperforms dense bi-encoders on multi-hop recall but still lags behind a strong BM25 baseline at open-corpus scale. The study demonstrates that graph-based retrieval offers interpretable multi-hop evidence paths and degrades more gracefully than dense retrieval as corpus size grows, but its advantage depends heavily on corpus scale and entity-index quality. Released with code and evaluation artifacts, Aethel provides a reproducible benchmark for multi-hop financial diligence.