Paste a document and see exactly how it splits for embeddings — chunk size, overlap, token estimate per chunk, and what your retriever actually indexes. Tokens are estimated (~4 chars/token); use it to reason about retrieval, not for exact billing.
Too big, and a chunk mixes topics — the embedding is muddy and retrieval pulls the wrong thing. Too small, and you lose the context that makes an answer correct. Overlap keeps a sentence that straddles a boundary findable from both sides. There's no universal best — measure it on your own questions. That's the whole point of an eval set.