If RAG performs poorly, the cause is usually not the model but how the documents were split. This step is the most decisive part of quality.
01. Chunk Size
The common range is 256 to 1024 tokens with some overlap between chunks. Too small loses context; too large drags irrelevant text into the answer.
02. The Parent-Child Approach
Current best practice: search with small chunks but send the model the larger section that chunk sits in. You get both precise retrieval and sufficient context.
03. Respect Structure
Rather than cutting blindly by character count, split on heading, clause and table boundaries. In contracts and procedure documents this single change lifts precision noticeably.
04. Tables and Annexes
Tables lose their meaning when flattened into plain text. They need to be interpreted row by row and rendered into text, or handled separately. Otherwise the system keeps getting numerical questions wrong.
05. The Embedding Model
For Turkish content the choice matters. Multilingual models generally perform well, but don't decide without testing on your own documents. If the model changes, the whole archive has to be reprocessed.
06. Metadata
Attach source document, section, date and access label to every chunk. Filtering, citation and authorisation all run through these fields.