The question.
Questions spanning many documents are difficult to answer with isolated chunks, but adding a knowledge graph introduces extraction errors and indexing cost.
The study tests where structured retrieval contributes beyond a well-tuned sparse-plus-dense baseline, instead of assuming graph retrieval is universally better.
The proposed approach.
- Preserve source offsets and document versions during ingestion
- Compare sparse, dense, reranked, and graph-assisted retrieval
- Use entity links only when supported by source evidence
- Check answer claims against the passages actually retrieved
Target architecture.
DocumentsHybrid indexEntity graphRetrieval plannerCited synthesisClaim checks
The difficult parts.
- Entity resolution and unsupported relationships
- Evaluation questions that favor one retrieval method
- Keeping citations valid after source updates
How to evaluate it.
- Retrieval recall and answer correctness by question type
- Citation support and unsupported-claim rate
- Indexing cost, query latency, and graph ablations
Planned deliverables.
- A retrieval-comparison benchmark
- A source-preserving research architecture
- A graph-versus-hybrid evaluation report
Design principles.
A citation must support the claim beside it.
Add structure when it improves the measured task.