ContextQMD
Back to Rtk

How RTK Savings Work

docs/guide/resources/savings-explained.md

0.44.03.3 KB
Original Source

How RTK Savings Work

RTK cuts up to 90% of the bash output your agent reads. This page explains what that number measures, what it does not measure, and how it reaches your bill.

What RTK filters

RTK sits between your agent and the CLI. When the agent runs a shell command, RTK executes it, compresses the output, and returns the compressed version.

agent runs a shell command
        |
        v
   RTK filters the output
        |
        v
  agent reads the result

The only thing RTK changes is the bytes a shell command sends back. Everything RTK reports as "savings" is measured on those bytes.

The savings chain

Cost
├─ Input tokens
│  ├─ Bash output           <- the only part RTK filters
│  ├─ Your prompt
│  ├─ System prompt
│  └─ Conversation history
└─ Output tokens            <- what the model writes

Those bytes are one contributor to input tokens, alongside your prompt, the system prompt, and conversation history. Input tokens are in turn only part of the bill, which also counts output tokens.

So the reduction dilutes at every step: a large cut in bash output produces a smaller cut in input tokens, and a smaller one again in cost. A command showing 90% fewer output bytes does not make your session 90% cheaper.

This is why RTK reports bash output reduction rather than a cost figure. Bash output is the part RTK controls; the rest depends on your prompt, your model, how much the agent writes back, and how much of the conversation is replayed on each call.

Why the token counts are estimates

rtk gain estimates tokens as bytes / 4:

rust
// src/core/tracking.rs
pub fn estimate_tokens(text: &str) -> usize {
    // ~4 chars per token on average
    (text.len() as f64 / 4.0).ceil() as usize
}

RTK ships no real tokenizer by design. Embedding one would cost startup time, and it would require a tokenizer per model, or a per-session model lookup, which RTK does not implement.

The consequence is worth understanding:

  • The percentage is reliable. The same estimator is applied to the raw output and the filtered output, so the ratio between them holds regardless of the estimator's absolute accuracy.
  • The absolute token counts are approximate. They will not match your provider's billing. Treat Input tokens: 45,230 as an order of magnitude, not an invoice line.

If you need exact counts, run the raw and filtered output through your model's own tokenizer.

How to read rtk gain

ColumnWhat it actually is
InputEstimated tokens from raw command output, bytes / 4
OutputEstimated tokens after filtering, bytes / 4
SavedInput minus Output, in estimated tokens
Save%Reduction in bash output bytes

Save% is the meaningful number. It is a byte ratio, and it is accurate as a ratio.

What RTK does not reduce

  • Output tokens. RTK never touches what the model writes.
  • Your prompt, the system prompt, or conversation history. These are input tokens RTK has no visibility into.
  • Commands with no matching filter. These pass through untouched and are tracked at 0% savings. See rtk gain --history.

See also