Token Saving

Reduce token usage by optimizing input content before it reaches the model.

How It Works

Request Arrives

JSON, HTML, Markdown, or text

Original input tokens

QuilrAI Compresses

JSON -> TOON

Markup -> clean text

Sent to LLM

Fewer prompt tokens

Meaning preserved ✓

QuilrAI

1. Request Arrives - Your app sends a normal API call
2. Gateway Optimizes - Enabled transforms rewrite eligible text fields to use fewer tokens
3. Forwarded to LLM - Optimized input is sent to the provider with no SDK changes

Compression Methods

Smart JSON Compression - Up to 20% savings

Converts eligible JSON objects or arrays in LLM inputs to TOON format - ideal for tool call responses and structured data.

Before	After
{"name": "John", "age": 30}	name:John|age:30

HTML to Text

Strips HTML tags and extracts clean text - removes markup overhead from scraped pages or rich content.

Before	After
<p class="intro"><b>Hello</b> world</p>	Hello world

Markdown to Text

Removes Markdown syntax characters that consume tokens without adding meaning for the LLM.

Before	After
## Hello **world**	Hello world

Text Compression

Compresses verbose plain text while preserving the original meaning. It removes low-value prose noise and separator noise while avoiding structured-looking lines.

Before	After
Please review the following statement and the context which was actually very repetitive.	Review statement and context.

Configuration Keys

Token saving is configured per API key. Enable only the transforms that match the traffic sent through that key.

{
  "smart_json_compression": false,
  "html_to_text": false,
  "markdown_to_text": false,
  "text_compression": false
}

Setting	Behavior
smart_json_compression	Converts eligible JSON objects or arrays to TOON when it reduces token usage.
html_to_text	Extracts clean text from HTML-heavy inputs.
markdown_to_text	Converts Markdown formatting to plain text and removes syntax-only tokens.
text_compression	Compresses verbose plain text while preserving meaning.

Input-Only Behavior

Compression is applied only to input tokens before they reach the LLM. Responses are returned untouched. Your application code stays exactly the same - no SDK changes, no prompt rewrites, just lower costs.