01 · Retrieval
RAG & GraphRAG
Understanding when graph structure improves evidence retrieval, multi-step reasoning and supported generation—and when simpler retrieval is the better engineering choice.
Research
Reliable retrieval for evidence-heavy decisions.
My current research interests centre on RAG/GraphRAG, trustworthy evidence retrieval, ESG document intelligence, data governance and task-aware AI systems.
02 · Responsible data
Traceability & Governance
Designing systems where evidence provenance, data quality, model routing, review and auditability are built into the decision pipeline.
03 · Applied domain
ESG & Public-interest Analytics
Applying data and AI to sustainability disclosure, green technology, educational equity and other evidence-intensive policy and business problems.
Current research · RAIDS Lab · University of Sydney
GraphRAG for Evidence-Based ESG Analysis
The project develops an end-to-end ESG report analysis system and benchmark to compare standard retrieval with multiple graph-based approaches under controlled conditions.
30real corporate ESG reports
2,015questions across four task levels
4graph-based retrieval methods
4major ESG frameworks organised
Methods: PaddleOCR, LLM-based normalisation, embeddings, hybrid retrieval, reranking, DualChannel, HippoRAG v1/v2, LightRAG and HyperGraphRAG. The evaluation considers factual retrieval, multi-step reasoning, summarisation, evidence-grounded generation, prompt token cost and graph size.
Main interpretation: GraphRAG should not replace standard RAG for every query. A practical architecture routes simple fact questions to lower-cost retrieval and sends complex, cross-evidence or high-risk questions to graph-based methods suited to the task.