Agentic GraphRAG for accounting regulatory research
Naïve RAG breaks on regulatory text — chunking cuts clauses in half, cosine ignores recency + authority, inter-document references aren't followed, and multi-hop questions get one-hop answers. The challenge: give retrieval a structure to walk. Entities as nodes, references / hierarchies / repeals / jurisdiction as typed edges. Then layer a small set of specialised agents (planner, extractor, verifier, clarifier) over the graph so the system can actually answer 'is X deductible for an exempt entity?' correctly. Real European regulation; starter code on Taxxa's GitHub.
Naïve RAG breaks on regulatory text — chunking cuts clauses in half, cosine ignores recency + authority, inter-document references aren't followed, and multi-hop questions get one-hop answers. The challenge: give retrieval a structure to walk. Entities as nodes, references / hierarchies / repeals / jurisdiction as typed edges. Then layer a small set of specialised agents (planner, extractor, verifier, clarifier) over the graph so the system can actually answer 'is X deductible for an exempt entity?' correctly. Real European regulation; starter code on Taxxa's GitHub.
Datasets, full brief and submission format live on the public Notion hub. Read it carefully before you start.
| Criterion | Max |
|---|---|
Retrieval quality Right nodes surfaced, multi-hop questions answered correctly. | 10 |
Graph design Entity + relation extraction holds up on regulation's real structure. | 10 |
Agent orchestration Planner / extractor / verifier / clarifier loop earns its keep on hard queries. | 10 |
Demo clarity Live demo makes the system's reasoning legible to a non-builder. | 5 |
The deadline is strict at Sun 15:00. Submissions go by email to the partner — clicking Submit via email opens a pre-filled draft from your client.