Structure-Aware Graph Retrieval for Long Annual Reports

This project studies structure-aware retrieval for evidence grounding over long annual reports. The key question is not only whether retrieved text is semantically similar, but whether the system can navigate document structure to find the exact evidence supporting an answer.

The method builds a typed metadata evidence graph over reports, pages, retrieval objects, entities, and metric categories. Graph expansion is used as a candidate-generation step, followed by cross-encoder reranking. The analysis shows that structure helps selectively: same-page and entity links can bridge evidence, same-metric links help modestly, and adjacent-page links can introduce distracting neighbors.

My contribution was to design the GraphRAG retrieval experiments, implement the typed graph expansion and edge ablations, evaluate object Recall@10, page Recall@10, MRR, and validation splits, and interpret the trade-off between relaxed page localization and exact evidence-object grounding.

• Constructed a typed evidence graph connecting reports, years, pages, objects, entities, and metric categories.
• Tested graph expansion as retrieval candidate generation rather than as final ranking.
• Ran edge-type ablations to separate useful structure from noisy proximity links.
• Added a lightweight structure-aware routing and ordering layer for interpretable retrieval control.
• Presented the work as under review, with enough method clarity for portfolio readers but without exposing the full manuscript.

Typed Evidence Graph

A high-level view of the retrieval design: annual reports are parsed into pages and evidence objects, linked through typed relations, expanded through graph neighborhoods, and reranked for object-level grounding.