llm-prompt-refiner 0.1.2

A lightweight Python library for optimizing and cleaning LLM inputs

Prompt Refiner

🧹 A lightweight Python library for optimizing and cleaning LLM inputs. Save 10-20% on API costs by removing invisible tokens, stripping HTML, and redacting PII.

⭐ If you find this useful, please star us on GitHub! ⭐

---

🎯 Perfect for:

RAG Applications • Chatbots • Document Processing • Production LLM Apps • Cost Optimization

---

Why use Prompt Refiner?

Stop paying for invisible tokens and dirty data.

Feature	Before (Dirty Input)	After (Refined)
HTML Cleaning	<div><b>Hello</b> world</div>	Hello world
Whitespace	User input\n\n\n here	User input here
PII Redaction	Call me at 555-0199	Call me at [PHONE]
Deduplication	Same text.\n\nSame text.\n\nDifferent.	Same text.\n\nDifferent.
Token Cost	❌ 150 Tokens	✅ 85 Tokens (Saved 43%)

📦 It's this easy:

from prompt_refiner import StripHTML, NormalizeWhitespace

cleaned = (StripHTML() | NormalizeWhitespace()).run(dirty_input)

✨ Key Features

• 🪶 Zero Dependencies - Lightweight core with no external dependencies
• ⚡ Blazing Fast - < 0.5ms per 1k tokens overhead, negligible impact on API latency
• 🔧 Modular Design - 5 focused modules: Cleaner, Compressor, Scrubber, Analyzer, Packer
• 🚀 Production Ready - Battle-tested operations with comprehensive test coverage
• 🎯 Type Safe - Full type hints for better IDE support and fewer bugs
• 📦 Easy to Use - Modern pipe operator syntax (|), compose operations like LEGO blocks

Overview

Prompt Refiner helps you clean and optimize prompts before sending them to LLM APIs. By removing unnecessary whitespace, duplicate characters, and other inefficiencies, you can:

• Reduce token usage and API costs
• Improve prompt quality and consistency
• Process inputs more efficiently

Status

This project is in early development. Features are being added iteratively.

Installation

# Basic installation (lightweight, zero dependencies)
pip install llm-prompt-refiner

# With precise token counting (optional, installs tiktoken)
pip install llm-prompt-refiner[token]

Installation Modes

• Default (Lightweight): Zero dependencies, uses character-based token estimation
• Precise Mode: Installs tiktoken for accurate token counting with no safety buffer

To use precise mode, pass a model parameter:

from prompt_refiner import CountTokens, ContextPacker

# Default: estimation mode (no model parameter)
counter = CountTokens()
packer = ContextPacker(max_tokens=1000)

# Opt-in: precise mode with tiktoken
counter = CountTokens(model="gpt-4")
packer = ContextPacker(max_tokens=1000, model="gpt-4")

Quick Start

from prompt_refiner import StripHTML, NormalizeWhitespace, TruncateTokens

# ✨ The Pythonic "Pipe" Syntax (Recommended)
pipeline = (
    StripHTML()
    | NormalizeWhitespace()
    | TruncateTokens(max_tokens=1000)
)

raw_input = "<div>  User input with <b>lots</b> of   spaces... </div>"
clean_prompt = pipeline.run(raw_input)
# Output: "User input with lots of spaces..."

Alternative: Fluent API

Prefer method chaining? Use the traditional fluent API:

from prompt_refiner import Refiner, StripHTML, NormalizeWhitespace, TruncateTokens

pipeline = (
    Refiner()
    .pipe(StripHTML())
    .pipe(NormalizeWhitespace())
    .pipe(TruncateTokens(max_tokens=1000))
)

clean_prompt = pipeline.run(raw_input)

💡 Why pipe operator? More concise, Pythonic, and familiar to LangChain/LangGraph users.

📊 Proven Effectiveness

We benchmarked Prompt Refiner on 30 real-world test cases (SQuAD + RAG scenarios) to measure token reduction and response quality:

Strategy	Token Reduction	Quality (Cosine)	Judge Approval	Overall Equivalent
Minimal	4.3%	0.987	86.7%	86.7%
Standard	4.8%	0.984	90.0%	86.7%
Aggressive	15.0%	0.964	80.0%	66.7%

Key Insights:

• Aggressive strategy achieves 3x more savings (15%) vs Minimal while maintaining 96.4% quality
• Individual RAG tests showed 17-74% token savings with aggressive strategy
• Deduplicate (Standard) shows minimal gains on typical RAG contexts
• TruncateTokens (Aggressive) provides the largest cost reduction for long contexts
• Trade-off: More aggressive = more savings but slightly lower judge approval

Example: RAG with duplicates

• Minimal (HTML + Whitespace): 17% reduction
• Standard (+ Deduplicate): 31% reduction
• Aggressive (+ Truncate 150 tokens): 49% reduction 🎉

💰 Cost Savings: At scale (1M tokens/month), 15% reduction saves ~$54/month on GPT-4 input tokens.

📖 See full benchmark: benchmark/custom/README.md

⚡ Performance & Latency

"What's the latency overhead?" - Negligible. Prompt Refiner adds < 0.5ms per 1k tokens of overhead.

Strategy	@ 1k tokens	@ 10k tokens	@ 50k tokens	Overhead per 1k tokens
Minimal (HTML + Whitespace)	0.05ms	0.48ms	2.39ms	0.05ms
Standard (+ Deduplicate)	0.26ms	2.47ms	12.27ms	0.25ms
Aggressive (+ Truncate)	0.26ms	2.46ms	12.38ms	0.25ms

Key Insights:

• ⚡ Minimal strategy: Only 0.05ms per 1k tokens (faster than a network packet)
• 🎯 Standard strategy: 0.25ms per 1k tokens - adds ~2.5ms to a 10k token prompt
• 📊 Context: Network + LLM TTFT is typically 600ms+, refining adds < 0.5% overhead
• 🚀 Individual operations (HTML, whitespace) are < 0.5ms per 1k tokens

Real-world impact:

10k token RAG context refining: ~2.5ms overhead
Network latency: ~100ms
LLM Processing (TTFT): ~500ms+
Total overhead: < 0.5% of request time

🔬 Run yourself: python benchmark/latency/benchmark.py (no API keys needed)

🎮 Interactive Demo

Try prompt-refiner in your browser - no installation required!

🚀 Launch Interactive Demo →

Play with different strategies, see real-time token savings, and find the perfect configuration for your use case. Features:

• 🎯 6 preset examples (e-commerce, support tickets, docs, RAG, etc.)
• ⚡ Quick strategy presets (Minimal, Standard, Aggressive)
• 💰 Real-time cost savings calculator
• 🔧 All 7 operations configurable
• 📊 Visual metrics dashboard

5 Core Modules

Prompt Refiner is organized into 5 specialized modules:

1. Cleaner - Clean Dirty Data

• StripHTML() - Remove HTML tags, convert to Markdown
• NormalizeWhitespace() - Collapse excessive whitespace
• FixUnicode() - Remove zero-width spaces and problematic Unicode

2. Compressor - Reduce Size

• TruncateTokens() - Smart truncation with sentence boundaries
  • Strategies: "head", "tail", "middle_out"
• Deduplicate() - Remove similar content (great for RAG)

3. Scrubber - Security & Privacy

• RedactPII() - Automatically redact emails, phones, IPs, credit cards, URLs, SSNs

4. Analyzer - Show Value

• CountTokens() - Track token savings and optimization impact
  • Estimation mode (default): Character-based approximation (1 token ≈ 4 chars)
  • Precise mode (with tiktoken): Exact token counts using OpenAI's tokenizer

5. Packer - Context Budget Management

• ContextPacker() - Intelligently pack items into token budgets with priority-based selection
  • Perfect for RAG applications and context window management
  • Priority constants: PRIORITY_SYSTEM, PRIORITY_USER, PRIORITY_HIGH, PRIORITY_MEDIUM, PRIORITY_LOW
  • Estimation mode: Applies 10% safety buffer to prevent context overflow
  • Precise mode: Uses 100% of token budget with accurate counting

Complete Example

from prompt_refiner import (
    # Cleaner
    StripHTML, NormalizeWhitespace, FixUnicode,
    # Compressor
    Deduplicate, TruncateTokens,
    # Scrubber
    RedactPII,
    # Analyzer
    CountTokens,
    # Packer
    ContextPacker, PRIORITY_SYSTEM, PRIORITY_USER, PRIORITY_HIGH
)

original_text = """<div>Your messy input here...</div>"""

# Create token counter to track savings
counter = CountTokens(original_text=original_text)

# Build the complete pipeline with all 4 modules
pipeline = (
    StripHTML(to_markdown=True)
    | NormalizeWhitespace()
    | FixUnicode()
    | Deduplicate(similarity_threshold=0.85)
    | TruncateTokens(max_tokens=500, strategy="head")
    | RedactPII(redact_types={"email", "phone"})
)

# Run and analyze
result = pipeline.run(original_text)
counter.process(result)

print(counter.format_stats())
# Output:
# Original: 8 tokens
# Cleaned: 5 tokens
# Saved: 3 tokens (37.5%)

Examples

Check out the examples/ folder for detailed examples organized by module:

• cleaner/ - HTML cleaning, whitespace normalization, Unicode fixing
• compressor/ - Smart truncation, deduplication
• scrubber/ - PII redaction
• analyzer/ - Token counting and cost savings
• packer/ - Context budget management with priorities for RAG
• all_modules_demo.py - Complete demonstration

Development

This project uses uv for dependency management and make for common tasks.

# Install dependencies
make install

# Run tests
make test

# Format code
make format

License

MIT