RAG chunking estimator

Reference scenarios

Calculated using the default assumptions above, varying the first input only.

How this estimate works

Each document is chunked independently. Chunks per document are one plus the rounded-up remaining token length divided by chunk size minus overlap, with a minimum of one for nonempty documents. Embedded token volume includes duplicated overlap but avoids padding the last chunk. Raw vector storage assumes float32 embeddings and excludes index and metadata overhead. Tokens per page and the embedding rate are illustrative inputs. Document parsing, headings, semantic boundaries and tokenizer differences affect actual results. Use a sample ingestion to tune these assumptions before a full run.

Start with your measured workload, adjust one assumption at a time, and compare the resulting capacity. Save the scenario link to return to these inputs or share it with your team. Reference scenarios remain fixed so you can compare against a consistent baseline.