also chunk size, text splitting
Splitting documents into the pieces that get embedded and retrieved. The boundaries decide what the model can see together, so chunk size is a decision about what a complete answer looks like, not a tuning knob.
You have done this if
You split policies by heading instead of every 512 tokens, after an answer quoted a rule without the exception in the next paragraph.
Say it in a review
We chunk on document structure and keep the heading path in each chunk, so a rule and its exception arrive together.
On the AI Application map Ingestion, Vector Index
Read Chunk boundaries decide what your RAG system can know · Chunk for the question, not for the token limit