Back to calculator
AI & AutomationGenerated accuracy model review: 2026-05-06
RAG Chunk Calculator accuracy and testing
This page documents how Calwarden builds trust for this calculator: source notes, regression-style default tests, limitations, assumptions, unit handling, rounding, and visible change history.
Formula source
- Total words = Documents x Average words per document. Estimated tokens = Total words / Average words per token. Effective chunk size = Chunk size - Overlap. Estimated chunks = ceiling((Estimated tokens - Overlap) / Effective chunk size).
Test cases
Default input regression check
Generated from the checked-in default inputs and calculate function.
Inputs
- Documents: 500
- Average Words per Document: 900
- Average Words per Token: 0.75
- Chunk Size: 800 tokens
- Overlap: 100 tokens
Expected outputs
- Estimated Chunks: 857
- Estimated Tokens: 600,000
- Effective Chunk Size: 700 tokens
- Average Chunks per Document: 1.71
Known limitations
- This page documents the checked-in calculator logic and does not certify results for every real-world edge case.
- Use the result as a planning estimate unless the calculator page states a stricter purpose.
Assumptions
- The estimate uses average document length rather than parsing each file.
- Tokenizer behavior varies by model and language.
- Overlap must be smaller than chunk size.
Unit handling
- Documents accepts numeric values.
- Average Words per Document accepts numeric values.
- Average Words per Token accepts numeric values.
- Chunk Size accepts values in tokens.
- Overlap accepts values in tokens.
- Effective Chunk Size is displayed in tokens.
Rounding rules
- Number displays use up to 2 decimal places in the calculator UI.
- Input precision follows configured step values where provided.
Version history
1.0.0 - 2026-05-06
Generated accuracy and testing page from the current rag-chunk-calculator calculator definition.
Change log
2026-05-06
Published default regression test documentation and metadata fallbacks for RAG Chunk Calculator.
Calculator transparency
Accuracy pages are generated from checked-in calculator definitions and can be upgraded with curated review metadata over time.