llm-sorter

A Python library that uses LLMs as comparators for semantic sorting. Sort any list by meaning, tone, complexity, urgency, or any criteria expressible in natural language. This library uses an OpenRouter API key to interface with LLM providers.

Installation

pip install llm-sorter

Requires Python 3.10+

Quick Start

from llm_sorter import LLMSorter

sorter = LLMSorter(api_key="your-openrouter-api-key")

# Sort by reading complexity
passages = [
    "The quantum entanglement phenomenon demonstrates non-local correlations...",
    "The cat sat on the mat.",
    "Climate change affects global weather patterns in complex ways."
]
result = sorter.sort(
    items=passages,
    prompt="Sort by reading level, simplest first"
)

# Sort support tickets by urgency
tickets = [
    "Login button color looks off",
    "CRITICAL: Production database is down, all users affected",
    "Would be nice to have dark mode"
]
urgent_first = sorter.sort(
    items=tickets,
    prompt="Sort by urgency, most critical first"
)

How It Works

llm-sorter implements merge sort with LLM-powered comparisons. Instead of numeric comparison, the model decides which of two items should come first based on your prompt.

This enables sorting by criteria that have no numeric representation: persuasiveness, professionalism, emotional intensity, policy compliance, or any semantic property. It can also enable the sorting of large lists that cannot be sorted within a single LLM call (with the tradeoff being significant more API calls, and the associated cost and latency of each call).

When to Use This

Good fit:

• Semantic/subjective ordering (urgency, quality, tone, reading level)
• No training data available — works zero-shot
• Rapid prototyping or internal tools
• Small to medium lists (n < 100-300)
• Approximate "human-like" rankings are acceptable

Examples:

• Sort writing samples by reading level
• Prioritize support tickets by urgency
• Rank answers by helpfulness
• Order reviews by sentiment strength

When NOT to Use This

Avoid when:

• You need strict determinism or reproducibility
• A simple numeric key exists (timestamps, severity levels)
• Very large n — O(n log n) API calls get expensive

Limitations

1. Non-transitive comparisons: LLMs can produce A > B, B > C, C > A cycles
2. Non-deterministic: Results may vary between runs
3. Cost at scale: 100 items ≈ 664 comparisons, 1000 items ≈ 10,000 comparisons
4. Prompt sensitivity: Small wording changes can affect results
5. Potential bias: Length, style, and positional biases exist

License

MIT