Meridian's charter-desk assistant needs to retrieve the exact clause a broker is asking about from a 40-page charterparty contract, without pulling in half of an unrelated clause. Which chunking approach best fits this source document?
Select an answer to reveal the explanation.
Short Explanation
Charterparty contracts are already organized into numbered clauses, so let the document tell you where to cut - chunk on clause boundaries and each retrieved piece is a complete, self-contained legal thought instead of half a sentence from clause 12 glued to half of clause 13.
Full Explanation
Structure-aware splitting on clause boundaries is correct because charterparty contracts already carry an explicit, machine-detectable structure - numbered clauses - and respecting that structure guarantees each chunk is a complete, coherent unit of meaning that can be retrieved and cited on its own. Fixed 500-character windows are wrong because they cut mid-clause with no awareness of the document's structure, frequently splitting one legal idea across two chunks or merging the tail of one clause with the head of the next, which corrupts both the embedding and the citation. Treating the whole 40-page contract as one chunk is wrong because it defeats the purpose of retrieval entirely - the embedding for a 40-page block is a blurry average of everything in it, so a query about demurrage will not score meaningfully higher against that chunk than a query about laytime, and the model receives far more irrelevant text than the question needs. Splitting one chunk per page is wrong for the same structural reason as fixed windows: clauses routinely run across a page break or share a page with an unrelated clause, so a page boundary is not a semantic boundary. The exam-relevant distinction is size-based versus structure-aware splitting, and legal/contract text is the canonical case for structure-aware.