prompt-bonsai 0.1.1

Intelligent prompt compression for LLMs — cut token costs without losing meaning

🌳 Prompt Bonsai

Cut your LLM token costs by 30–70% without losing meaning.

Prompt Bonsai intelligently compresses prompts for Large Language Models (LLMs), reducing token usage and API costs while preserving semantic meaning and output quality.

Why Prompt Bonsai?

Problem	Solution
Token costs eating your budget?	30–70% token reduction with quality guarantees
Context window too small?	Fit more content in limited context
Prompts full of filler?	Remove semantic redundancy automatically
JSON/code bloating tokens?	Structural optimization preserves syntax
Worried about quality loss?	Built-in quality assessment with configurable thresholds

Installation

pip install prompt-bonsai

With optional dependencies:

pip install prompt-bonsai[openai,langchain]  # For integrations
pip install prompt-bonsai[all]                  # Everything

Quick Start

One-line compression

from prompt_bonsai import compress

long_prompt = "I would like to kindly request your assistance with..." * 10
short = compress(long_prompt, ratio=0.5)

print(f"Saved {len(long_prompt) - len(short)} characters")

Full control with Compressor

from prompt_bonsai import Compressor

compressor = Compressor(
    strategy="hybrid",      # "semantic", "structural", or "hybrid"
    target_ratio=0.4,        # Aim for 40% reduction
    min_quality=0.90,       # Reject if quality drops below 90%
    preserve=["{user_id}"], # Keep template variables
    model_name="gpt-4",     # Optimize for specific tokenizer
    verbose=True,           # Show compression details
)

result = compressor.compress(your_prompt)

print(result.text)                          # Compressed prompt
print(result.original_tokens)             # Before
print(result.compressed_tokens)           # After
print(result.quality_report.overall_score)  # Quality: 0.0–1.0

Compression Strategies

Strategy	Best For	How It Works
Semantic	Conversational prompts	Removes filler words, redundant phrases, weak adverbs
Structural	Code, JSON, XML	Normalizes whitespace, minimizes structured data, removes comments
Hybrid	Mixed content (default)	Auto-detects prompt type and applies optimal strategy

Real-World Example

from prompt_bonsai import Compressor

compressor = Compressor(target_ratio=0.5, verbose=True)

system_prompt = """
You are an expert Python developer. Please review the following code
and suggest improvements for readability, performance, and
maintainability. Consider PEP 8 compliance, type hints, documentation,
error handling, and testing coverage in your analysis.
"""

result = compressor.compress(system_prompt)

Output:

┏━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Metric              ┃ Value                           ┃
┡━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ Strategy            │ HybridCompressor                │
│ Original Tokens     │ 67                              │
│ Compressed Tokens   │ 34                              │
│ Tokens Saved        │ 33 (49.3%)                      │
│ Quality Score       │ 0.912                           │
│ Semantic Similarity │ 0.845                           │
│ Structural Integrity│ 1.000                           │
└─────────────────────┴─────────────────────────────────┘

Integration Examples

With OpenAI

from openai import OpenAI
from prompt_bonsai import compress

client = OpenAI()

prompt = "Your very long prompt here..."
compressed = compress(prompt, ratio=0.4)

response = client.chat.completions.create(
    model="gpt-4",
    messages=[{"role": "user", "content": compressed}]
)

With LangChain

from prompt_bonsai import Compressor

compressor = Compressor(target_ratio=0.3)

# Compress before creating your prompt
compressed_context = compressor.compress(large_document).text

# Use in your LangChain pipeline
prompt = PromptTemplate.from_template(
    "Answer based on: {context}\n\nQuestion: {question}"
)

Configuration

from prompt_bonsai.config import CompressionConfig, CompressionStrategy

config = CompressionConfig(
    strategy=CompressionStrategy.HYBRID,
    target_ratio=0.5,
    min_quality=0.90,
    max_iterations=3,
    preserve_patterns=["{variable}", "{{template}}"],
    tokenizer_backend="auto",  # "tiktoken", "huggingface", or "auto"
    model_name="gpt-4",        # For accurate token counting
)

compressor = Compressor(config=config)

CLI Usage

# Compress a file
prompt-bonsai compress input.txt --ratio 0.5 --output compressed.txt

# Compress from stdin
echo "Your long prompt..." | prompt-bonsai compress --ratio 0.4

# Check token count
prompt-bonsai tokens "Your prompt here" --model gpt-4

Benchmarks

Run benchmarks against standard datasets:

python benchmarks/run_benchmarks.py

API Reference

See full documentation.

Contributing

We welcome contributions! See CONTRIBUTING.md.

git clone https://github.com/vodalachakshu123/prompt-bonsai.git
cd prompt-bonsai
pip install -e ".[dev]"
pytest

License

MIT License — see LICENSE.