How Should Urdu Documents Be Chunked for Retrieval? A Budget-Matched Study of Chunk Size, Overlap, Sentence Boundaries and Tokenization

This preprint studies how Urdu documents should be chunked for retrieval. It builds a parallelUrdu-English corpus from UQA (machine-translated SQuAD2.0) and the original SQuAD2.0 articles:150 articles, 7,362 test questions, and 1,007,122 Urdu words. The study compares five chunkers (fixed words, fixed tokens, sentence-aware, recursive-English,and recursive-Urdu), four multilingual encoders (bge-m3, multilingual-e5-large-instruct,Qwen3-Embedding-0.6B, and LaBSE), BM25 whitespace and character n-grams, and BM25+dense reciprocalrank fusion. The main grid contains 360 runs: word sizes 50-800, token budgets 64-256, and overlap0/10/25%. A separate four-run bge-m3 check extends the word grid down to 10/25/40/50. Results. Among the tested 50/100/200-word chunks, 50 words is the best tested size for all fourdense encoders. The bge-m3 lower-size check is monotonic (0.418 at 10, 0.689 at 25, 0.756 at 40,0.773 at 50), so 50 remains the best tested size; the 40/50 intervals overlap. Token argmaxes differ(64/128/256) because tokenizer fertility differs by about 2.3x (1.34 vs. 3.07 tokens per Urdu word).Urdu-aware sentence splitting beats the English recursive default by 5.5-6.9 points at 50 words andreverses at 200 words. 25% overlap is associated with 2.6-7.2 more points on this corpus.Orthographic normalization is a null on this already-normalized machine-translated corpus. Englishis easier on the parallel control but selects the same best tested fixed-word size. This is a preprint and has not been peer reviewed. The corpus is machine-translated andsingle-domain; no native-authored evaluation set, human relevance/context judgments, or externalpublished baseline was reproduced. The lower-size check covers bge-m3 only. Code, configurations,question IDs, and per-question metrics are available at:https://github.com/devxyasir/urdu-chunking-strategy UQA has no explicit licence; the artifact excludes UQA text. The paper's Figure 1 contains a shortillustrative quotation.