Compress

Compress

Text compression and token optimization for efficient LLM communication

Category: development Source: JuliusBrussee/caveman

Compress: Text Compression and Token Optimization for Efficient LLM Communication

What Is This

The Compress skill for the Happycapy Skills platform provides developers with a powerful tool for text compression and token optimization. It is specifically designed to reduce the token count of text inputs and outputs when interacting with large language models (LLMs). By leveraging state-of-the-art compression algorithms and encoding strategies, Compress transforms verbose or repetitive text into a more compact form, thereby enhancing the efficiency of LLM-based applications. The skill integrates seamlessly with the Happycapy platform and is based on the open source implementation at caveman-compress.

Compress supports both lossless and lossy compression methods, making it suitable for scenarios where either fidelity or maximum compactness is required. The skill can be used as a standalone text compressor or as part of a broader LLM communication pipeline, especially when token budgets are tight or API costs need to be minimized.

Why Use It

Large language models are typically billed and rate-limited based on token usage. Reducing token count without losing essential information can significantly decrease costs, speed up interactions, and enable longer context windows. Compress addresses these challenges by:

  • Shrinking text length using advanced compression techniques.
  • Optimizing tokenization for popular LLMs such as GPT-3/4, Claude, and others.
  • Preserving information integrity (when using lossless mode), or maximizing compactness (in lossy mode) for scenarios where some loss of detail is acceptable.
  • Supporting both compression and decompression, ensuring that compressed text can be restored to its original form when needed.

Use cases include compressing system prompts, user inputs, intermediate chain-of-thought steps, or storing large datasets as compressed text for LLM retrieval.

How to Use It

The Compress skill can be invoked via the Happycapy Skills platform, either through API calls or as part of a skill chain. Below are the typical usage patterns:

Basic Compression Example

from happycapy_skills import compress

## Compress a block of text (lossless mode)
input_text = "This is a very long and repetitive message. This is a very long and repetitive message."
compressed = compress.compress_text(input_text, mode="lossless")
print("Compressed:", compressed)

Basic Decompression Example

## Decompress the previously compressed text
decompressed = compress.decompress_text(compressed)
print("Decompressed:", decompressed)

Lossy Compression

If you need greater compression and can tolerate some information loss, use the lossy mode:

compressed_lossy = compress.compress_text(input_text, mode="lossy")
print("Lossy Compressed:", compressed_lossy)

Skill Configuration

You can configure the compression parameters via skill settings, such as:

  • mode: "lossless" (default) or "lossy"
  • target_tokens: Maximum target token count for the compressed text
  • model: Target LLM model or tokenizer, for optimal tokenization

Example:

compressed = compress.compress_text(
    input_text,
    mode="lossless",
    target_tokens=100,
    model="gpt-3.5-turbo"
)

Integration in Skill Chains

Compress can be combined with other skills in a Happycapy chain to minimize token usage at each step, especially for memory-intensive workflows.

When to Use It

Compress is especially valuable when:

  • You are working with LLMs that have strict input/output token limits.
  • Reducing API costs by minimizing token usage is a priority.
  • You need to transmit or store large texts efficiently within a prompt or conversation.
  • Your workflow involves intermediate representations that must fit within an LLM context window.
  • You want to maximize the number of messages or data packets exchanged with an LLM per session.

Typical scenarios include semantic search pipelines, multi-turn chat agents, and any application that operates near token limits.

Important Notes

  • Lossless vs Lossy: Use lossless mode for scenarios where information fidelity is critical, such as system instructions, code, or legal documents. Use lossy mode for summaries or data where minor loss is acceptable for higher compression.
  • Tokenization Awareness: The Compress skill can optimize for different LLM tokenizers. Always specify the model for best results.
  • Decompression Required: Compressed output is not directly readable by humans or LLMs without decompression. Always decompress before displaying or further processing.
  • Compatibility: The skill depends on the caveman-compress implementation. Ensure compatibility with your Happycapy and LLM environment.
  • Performance: Compression and decompression involve additional processing. Benchmark performance for high-throughput applications.
  • Security and Privacy: Compressed data may look opaque but is not encrypted. Sensitive information should still be handled according to best security practices.

For more information and advanced usage, refer to the caveman-compress documentation and the skill's SKILL.md.