Worked on the infiniflow/ragflow repository to deliver metadata-driven enhancements for large document retrieval and navigation. Focused on backend development and data processing, the work introduced RAPTOR layer metadata persistence and PDF bookmark outline storage, enabling targeted retrieval by abstraction level and richer navigation within documents. Leveraging Python and Elasticsearch, the implementation exposed new metadata fields for layer-based filtering and improved recall quality for definition queries. All changes were backward-compatible, requiring no schema modifications, and supported downstream entity extraction and search relevance. The approach emphasized additive, non-breaking updates, with careful integration of API and JSON-based data handling throughout the codebase.