AI LAB / RAG / Chunking
Chunk Size vs Answer Quality
Testing how document chunk size changes retrieval recall, context precision, and citation quality.
RAGChunkingEvaluation
- EXPERIMENT
- EXP-019
- STATUS
- complete
- DATE
- Jul 30, 2026
- TAGS
- RAG · Chunking · Evaluation
EXPERIMENT / HYPOTHESIS
Structure-aware chunks will beat a fixed token window across mixed technical documents.
- 01Create fixed chunks at three sizes
- 02Create heading-aware semantic chunks
- 03Evaluate retrieval and citation completeness
- Best recall
- 94.1%
- Citation precision
- 90.7%
- Documents
- 180
CONCLUSIONStructure-aware chunks had the best balance. Very small chunks improved retrieval precision but often lost the context needed for complete answers.
Variants
The test compared 256, 512, and 1,024-token windows with a structure-aware parser preserving headings and lists.
- Same embeddings
- Same retriever
- Same answer model
Finding
Chunking quality depended on document structure more than one globally optimal token count.
- Preserve semantic boundaries
- Include parent headings
- Measure citation completeness