Chunking (découpage de documents)
The operation that splits a document into pieces before embedding them for a RAG. Piece size and overlap decide what the model will be able to retrieve: too large and search returns noise, too small and the sentence loses the context that made it understandable. A piece that starts halfway through a table or cuts a definition in two will never be rescued by downstream reranking.
Strengths
- The cheapest quality lever: no model swap, no extra line in the prompt
- Explains in one sentence why a RAG answers off-target even though the data is properly indexed
Limitations
- No universal setting: the right size depends on document type and on the question asked
Best for
- Diagnosing a RAG that returns noise before touching the model