| name | tune-chunking |
| description | Tune chunk size, overlap, and structure-awareness against the eval and the context budget. Reach for this on a chunking question. |
Skill: Tune chunking
A chunk that splits the answer is a retrieval bug introduced at ingestion (§3 #2).
Step 1 — Inspect the failing chunks
Are answering units (tables, sections) split or buried? (§3 #2)
Step 2 — Adjust size + overlap
Tune against recall@k, not a guessed default (§3 #2).
Step 3 — Add structure-awareness
Keep tables/sections intact; add metadata for pre-filtering (§3 #2).
Step 4 — Check the context budget
Size × top-k + overhead vs window via ai_rag_engineering_calc.py chunk-budget (§3 #5).
Output
A chunking design tied to recall@k and the context budget. Traverse Tree 2 in the decision-trees file.