Quality Assessment

Prompt Bonsai evaluates compressed prompts on four dimensions:

Quality Metrics

Semantic Similarity (35% weight)

Measures word overlap between original and compressed text using Jaccard similarity.

Structural Integrity (25% weight)

Checks for broken brackets, quotes, and formatting. - Balanced parentheses (), [], {}, <> - Balanced quotes " - Preserved code structure

Information Retention (25% weight)

• Preserves capitalized terms and technical keywords
• Tracks preserve_patterns retention
• Measures keyword coverage

Readability Score (15% weight)

• Ideal sentence length: 15-25 words
• Penalizes overly long sentences
• Rewards clear structure

QualityReport

from prompt_bonsai import Compressor

compressor = Compressor(min_quality=0.90)
result = compressor.compress(prompt)

report = result.quality_report
print(f"Overall: {report.overall_score:.2f}")
print(f"Semantic: {report.semantic_similarity:.2f}")
print(f"Structural: {report.structural_integrity:.2f}")
print(f"Information: {report.information_retention:.2f}")
print(f"Readability: {report.readability_score:.2f}")

if report.warnings:
    print("Warnings:", report.warnings)

Handling Low Quality

If compression falls below min_quality, a QualityError is raised:

from prompt_bonsai import Compressor
from prompt_bonsai.exceptions import QualityError

compressor = Compressor(min_quality=0.95)

try:
    result = compressor.compress(prompt, target_ratio=0.8)
except QualityError as e:
    print(f"Compression too aggressive: {e}")
    # Fall back to gentler compression
    result = compressor.compress(prompt, target_ratio=0.4)