INDIAResearch
arXiv: StrucChunk doubles LegalBench-RAG recall by splitting legal texts along their structure
Three independent researchers propose StrucChunk, a retrieval method that splits legal documents at provision boundaries, resolves cross-references before embedding and combines keyword and semantic search. On LegalBench-RAG's 6,858 query-passage pairs it reached 48.18% character recall in the top eight results, against 23.59% for the best baseline. Tests covered five corpora, including Indian, EU and Thai law, in two languages.
Key points
- On 501 queries on India's criminal procedure code, recall in the top five rose from 0.557 to 0.804, mostly due to chunking.arXiv
- It kept 97.6% of documented cross-references intact, while recursive splitting preserved none.arXiv
- On 501 GDPR queries it reached top-five recall of 0.868, and it covered every GDPR section with 40.3% fewer chunks than recursive splitting.arXiv
- Fixed-size chunking could not represent 29.6% of Thai legal sections whose text exceeded the chunk limit, the paper says.arXiv
- The code and 5,256 evaluation queries across the five corpora are publicly released.arXiv