A GenAI engineer notices the charter-desk assistant sometimes answers demurrage questions using text from a laytime clause instead, and sometimes truncates a valid answer mid-sentence. The chunk size is currently very small (roughly 100 characters). What is the most likely fix?
Select an answer to reveal the explanation.
Short Explanation
Chunks that tiny are basically confetti - too small to hold a whole thought, so the model grabs a scrap of the wrong clause and then runs out of room mid-sentence on the right one; give each chunk room to hold a complete idea and both symptoms usually clear up.
Full Explanation
Increasing the chunk size is correct because both symptoms described - retrieving the wrong clause and truncating a valid answer mid-sentence - are classic signatures of chunks that are too small to capture a complete, coherent unit of meaning; at roughly 100 characters, a chunk may not even contain a full sentence, so its embedding is a noisy, ambiguous signal that can score closer to an unrelated clause than to the correct one, and the retrieved fragment may end before the relevant text does. Switching to a smaller embedding model with fewer dimensions addresses the wrong variable entirely - dimensionality affects embedding representational capacity and storage cost, not how much source text was captured in the chunk to begin with, and it would not resolve either symptom. Disabling chunking and embedding each contract as a single vector overcorrects in the other direction, producing the diluted, low-precision retrieval discussed elsewhere in this domain - a 40-page contract as one vector cannot discriminate between its own clauses. 'Unscored items returned by the retriever' is not a real retrieval-tuning concept - unscored items refer to exam-delivery mechanics, not a Vector Search parameter, making that option a distractor with no bearing on chunk quality. The exam-relevant skill is diagnosing retrieval symptoms back to the chunking-parameter root cause rather than reaching for an unrelated lever.