Plan the token demand of a retrieval augmented generation workload before estimating cost or capacity. Enter the fixed prompt, retrieved context, expected answer length, request volume, planning period, and a safety buffer. The calculator separates retrieval tokens from other input and output tokens, then reports both the base workload and a buffered budget so you can size quotas without hiding the assumptions.
The main result is the total buffered token allowance for the selected period. Retrieval share shows how much of the input is attributable to inserted passages, while tokens per request is useful for model-context and throughput planning.
Token counts depend on the selected model tokenizer. Use measured production averages when available; this is a planning estimate, not a billing invoice.
For 10,000 requests/day over 30 days, with 800 fixed input tokens, 5 chunks × 300 tokens, 500 output tokens, and a 15% buffer:
No. This calculator covers tokens sent to and generated by the language model. Add embedding workload separately if your provider bills it.
They equal the average retrieved chunks per request multiplied by the average tokens in each inserted chunk.
Include reranker tokens separately if it uses a billed language model; this page models the final generation request only.
Use a measured average for baseline planning and increase the safety buffer or run a high-volume scenario for peaks.
The period total can, but the per-request total must fit the model context limit after accounting for any provider-specific overhead.
| Variable | Meaning | Unit |
|---|---|---|
| R | Requests per day | requests/day |
| F | Fixed non-retrieved input | tokens/request |
| C × T | Retrieved chunks × tokens/chunk | tokens/request |
| O | Generated output | tokens/request |
| B | Safety buffer | percent |