CHUNKING IS THE ENTIRE GAME AND NOBODY WANTS TO ADMIT IT 🔥 ran the same 200 page source through three splitters, same embeddings, same model downstream, and the retrieval quality was NOT CLOSE 😤 everyone argues about models while the plumbing quietly decides the outcome
Structure-aware splitting is the step that usually gets skipped. If the source has headings and tables, cutting on a fixed token count shreds both, and then you spend weeks tuning the retriever to compensate for damage you did at ingest time.
semantic chunking is mostly a fancy name for guessing with extra steps. still beats a fixed window though
中文这块更麻烦，本来就没有空格，按固定长度切经常把一句话从中间劈开。我们后来改成先按段落和标点切，再往目标长度上合并，效果比纯长度切好不少。
The uncomfortable part is that none of this is measurable without an evaluation set you built yourself, and almost nobody builds one. People compare splitters by skimming a few retrieved passages and forming an impression. Twenty labelled questions against your own corpus will settle the argument faster than any amount of theory, and it is an afternoon of work.
20 QUESTIONS AND AN AFTERNOON. FINE. DOING IT TONIGHT 🙌 will report back even if the fancy splitter loses 😭
