prompt-compressor

Rule-based LLM prompt compression — 30+ algorithms, zero dependencies.

Compress prompts before sending them to any LLM API. No model required, no API calls, zero overhead. Works with OpenAI, Anthropic, AWS Bedrock, Cohere, or any token-based API.

🌐 Web app: prompt-compressor.com
📦 npm: prompt-compressor

Install

pip install prompt-compressor

Python API

from prompt_compressor import compress

result = compress(system_prompt, level="low")

print(result.output)           # compressed text
print(result.tokens_before)    # e.g. 1200
print(result.tokens_after)     # e.g. 820
print(result.saved)            # 380
print(result.reduction_pct)    # 31.7

Convenience function

from prompt_compressor import compress_text

compressed = compress_text(my_prompt, level="very-low")

With AWS Bedrock

import boto3
from prompt_compressor import compress

bedrock = boto3.client("bedrock-runtime", region_name="us-east-1")

result = compress(system_prompt, level="low")

response = bedrock.invoke_model(
    modelId="anthropic.claude-3-5-sonnet-20241022-v2:0",
    body=json.dumps({
        "anthropic_version": "bedrock-2023-05-31",
        "system": result.output,   # compressed!
        "messages": [{"role": "user", "content": user_message}],
        "max_tokens": 1024,
    }),
)

CLI

# Compress a file
prompt-compressor system_prompt.txt --level low

# With stats
prompt-compressor system_prompt.txt --level medium --stats

# Pipe
cat my_prompt.txt | prompt-compressor --level very-low

# Save to file
prompt-compressor input.txt --level low --output compressed.txt

Levels

Variable Protection

Variables in {{var}}, {var}, <tag>, and ${var} format are always protected by default:

result = compress("Hello {{name}}, query: {user_query}", level="medium")
# {{name}} and {user_query} are preserved exactly

Disable with protect_vars=False.

Algorithm Reference

Structural (none / very-low)

Language (very-low / low)

Code (low)

Deduplication (low / medium)

Extractive (medium)

Zero Dependencies

prompt-compressor uses only the Python standard library. No numpy, nltk, spacy, tiktoken, or anything else to install.

License

MIT © Tom Ketter