How does input compression handle code snippets in prompts?

All fenced code blocks, inline code, and JSON objects are detected and protected by Safe Vault, ensuring zero modification to executable code.

Context Window Optimization

High-speed input prompt context compression

Strip redundant semantic tokens, whitespace bloat, and RAG document noise before payloads reach your frontier model endpoint.

Average Input Savings

50% - 68%

Measured across RAG context, agent tool calls, and developer prompts.

Throughput Speed

< 2ms

In-memory deterministic compression with zero external network hops.

AST Syntax Drift

0.00%

Bit-for-bit guarantees on all isolated code and structured blocks.

The production engineering challenge

Large RAG documents, multi-turn chat histories, and massive agent tool schemas consume tens of thousands of tokens per call. Due to BPE sub-token splitting, repetitive formatting characters and word structures inflate context windows and multiply API costs.

Technical architecture & mechanics

Sub-token whitespace and delimiter normalization

Normalizes repetitive spaces, indentations, and punctuation clusters that cause modern BPE tokenizers to generate fragmented sub-tokens.

Semantic entropy pruning

Identifies and eliminates low-information words and conversational filler from user and system instructions while retaining key entities and directives.

Safe Vault code isolation

Locks source code, JSON schemas, SQL queries, and variable placeholders bit-for-bit so they bypass text compression completely.

Direct Input Compression API

const res = await fetch('https://api.curtly.dev/v1/compress', { method: 'POST', headers: { 'Authorization': 'Bearer ' + process.env.CURTLY_API_KEY, 'Content-Type': 'application/json', }, body: JSON.stringify({ prompt: verboseContextDocument, mode: 'balanced', // conservative | balanced | aggressive model: 'claude-fable-5-1', }), }); const { compressedPrompt, tokensSaved, savingsPercent } = await res.json();

Architecture comparison

Evaluation metric	Curtly.dev gateway	Traditional / naive approach
Compression Speed	< 2ms in-memory execution	150-400ms via secondary LLM summarizers
Code & JSON Safety	Safe Vault isolation ensures 0% corruption	Frequent truncation of quotes, brackets, and variables
Cost Overhead	Included in lightweight flat plan	Requires paying for extra summarizer model tokens

Frequently asked questions

No. Curtly applies deterministic token normalization and entropy reduction that preserves semantic meaning, instruction priority, and structural constraints.

Explore related platform features

Deterministic Syntax Shield

Ready to optimize your AI infrastructure?

Start optimizing in 60 seconds with a 1-line baseURL change. Free tier available with zero credit card required.

Start optimizing free Simulate token ROI