GitHub - metawake/prompt_compressor: Compresses LLM prompts while preserving semantic meaning to reduce token usage and cost.

Semantic Prompt Compressor

A Python library for compressing prompts while preserving their semantic meaning. This tool helps reduce token count for LLM prompts without losing important information.

Features

Modular, rule-based compression
Multiple compression profiles (safe, quality)
Configurable compression rules
Token count reduction analysis
Command-line interface for easy use

Installation

Usage

Command Line Interface

The package provides a CLI with two main commands: compress and analyze.

Basic Compression

prompt-compress compress "I am really interested in learning more about Python programming"

Output:

I'm interested in learning about Python programming

Show Compression Statistics

prompt-compress compress "I am really interested in learning more about Python programming" -s

Output:

=== Compression Statistics ===
Original tokens: 12
Compressed tokens: 9
Tokens saved: 3
Compression ratio: 25.00%

Detailed Analysis

prompt-compress compress "I am really interested in learning more about Python programming" -v

Output:

=== Compression Statistics ===
Original tokens: 12
Compressed tokens: 9
Tokens saved: 3
Compression ratio: 25.00%

=== Detailed Analysis ===
Original text: I am really interested in learning more about Python programming
Compressed text: I'm interested in learning about Python programming

Analyze Compression Opportunities

prompt-compress analyze "I am really interested in learning more about Python programming"

Output:

=== Compression Statistics ===
Original tokens: 12
Compressed tokens: 9
Tokens saved: 3
Compression ratio: 25.00%

Save Results to File

prompt-compress compress "I am really interested in learning more about Python programming" -o results.json

Python API

from prompt_compressor import PromptCompressor

# Initialize with default rules
compressor = PromptCompressor()

# Compress a prompt
compressed = compressor.compress("I am really interested in learning more about Python programming")

# Analyze compression
analysis = compressor.analyze_prompt("I am really interested in learning more about Python programming")
print(f"Original tokens: {analysis['original_tokens']}")
print(f"Compressed tokens: {analysis['compressed_tokens']}")
print(f"Compression ratio: {analysis['compression_ratio']:.2f}%")

Compression Rules

The compression rules are defined in YAML format and can be customized. The default rules include:

Remove fillers and greetings
Strip unnecessary modifiers
Collapse redundant phrases
Preserve technical terms and entities
Handle contractions properly

Example rules configuration:

rule_groups:
  remove_fillers:
    enabled: true
    patterns:
      - pattern: "I am really"
        replacement: "I'm"
      - pattern: "more about"
        replacement: "about"

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the MIT License - see the LICENSE file for details.