Chunk Size
Properties
tags
gen_aigen_ai/rag
created
08.02.2025, 11:56
modified
02.08.2026, 10:19
published
Empty
topics
Chunking
authors
Jakub
ai-assisted
No
Chunk size significantly impacts performance. Larger chunks provide more context, enhancing comprehension but increasing process time. Smaller chunks improve retrieval recall and reduce time but may lack sufficient context.
Generally, the chunk strategy and the size have to be fine tuned for specific situation, but the paper Searching for Best Practices in RAG provides at least some guide what should be good starting chunking size of 256 with overlap of 20 tokens:
| Chunk Size | Average Faithfulness | Average Relevancy |
|---|---|---|
| 2048 | 80.37 | 91.11 |
| 1024 | 94.26 | 95.56 |
| 512 | 97.59 | 97.41 |
| 256 | 97.22 | 97.78 |
| 128 | 95.74 | 97.22 |
| Table 3: Comparison of different chunk sizes for lyft_2021 dataset. |