HEMLOCK - Hybrid Entity Multi-hop Linkage Over Co-occurrence Knowledge

A multi-hop question needs documents it reaches only through entities it never names, and on $19$-$51$% of questions ColBERTv2's top $30$ misses one. Closing that gap usually costs a generative call per indexed passage or reasoning step. HEMLOCK makes none: a bipartite graph links text segments to the entities they mention, and a Personalised PageRank walk seeded from ColBERTv2 reaches bridge segments through shared entities for a cross-encoder to rerank. It spends zero LLM tokens at index and query time, serves a query in $96$\ms and needs no per-corpus tuning. On HotpotQA, 2WikiMultihopQA and MuSiQue the walk adds $4.2$-$5.4$pp of passage recall@$10$ over the same pipeline without it; no other system we evaluate that makes no generative calls retrieves more, LinearRAG at its published configuration included. HEMLOCK reaches $85$-$95$% of the document recall@$10$ of HippoRAG 2, whose graph is LLM-built, and exceeds IRCoT's by $14.0$-$16.3$pp. Where ColBERTv2's top-$30$ misses a gold document, no reranker over that list can recover it; the walk lifts $49.5$, $39.9$ and $21.7$% of them into the top ten, raising exact match under a frozen reader.