Semantic RBO

Consensus-building library that merges multiple stakeholders' prioritized lists using semantic clustering and Rank-Biased Overlap (RBO).

Features

• Semantic Clustering - Groups similar ideas using embedding vectors (MiniLM) and cosine similarity
• Rank-Biased Overlap - Weights rankings by position importance (top items matter more)
• Paraphrase Detection - Automatically recognizes when different people express the same idea differently
• Agreement Metrics - Quantifies how much stakeholders agree with each other
• Markdown Reports - Generates publication-ready consensus reports

Advanced Features

• Configurable Decay Functions - Beyond exponential decay, supports linear, logarithmic, square root, plateau, and custom decay functions for rank weighting
• Faction Detection - Automatically clusters submitters into factions based on agreement patterns, identifying coalitions and outliers
• Threshold Sensitivity Analysis - Test consensus stability across multiple similarity thresholds to validate results

Installation

npm install semantic-rbo

Requires Node.js 22.0.0 or higher.

Quick Start

import { buildConsensus } from 'semantic-rbo/builder';

const result = await buildConsensus({
    documents: [
        {
            docId: `alice`,
            steps: [
                `Improve user onboarding`,
                `Add dark mode support`,
                `Fix mobile responsiveness`
            ]
        },
        {
            docId: `bob`,
            steps: [
                `Better onboarding experience`,  // Similar to Alice's #1
                `Performance optimization`,
                `Dark theme option`               // Similar to Alice's #2
            ]
        }
    ]
});

console.log(result.consensus);
// Outputs ranked list with semantic grouping and agreement scores

Documentation

See the full documentation for:

• Getting Started Guide
• RBO Algorithm Explained
• Semantic Clustering
• Decay Functions
• Faction Analysis
• Core API Reference
• Builder API Reference

How It Works

The Pipeline

flowchart LR
    subgraph Input
        A[Alice's Priorities]
        B[Bob's Priorities]
        C[Carol's Priorities]
    end

    subgraph Process
        D[Understand<br/>Meaning]
        E[Group Similar<br/>Ideas]
        F[Merge<br/>Rankings]
    end

    subgraph Output
        G[Consensus<br/>List]
        H[Agreement<br/>Metrics]
    end

    A --> D
    B --> D
    C --> D
    D --> E --> F --> G
    F --> H

A Concrete Example

Here's how two stakeholders' priorities flow through the system:

flowchart TD
    subgraph "Step 1: Raw Input"
        A1["Alice: 1. Improve onboarding<br/>2. Add dark mode<br/>3. Fix mobile"]
        B1["Bob: 1. Better onboarding<br/>2. Performance<br/>3. Dark theme"]
    end

    subgraph "Step 2: Understand Meaning"
        E["Each item converted to<br/>a numeric meaning vector"]
    end

    subgraph "Step 3: Find Matches"
        M1["'Improve onboarding' ≈ 'Better onboarding'"]
        M2["'Add dark mode' ≈ 'Dark theme'"]
        M3["'Fix mobile' — unique"]
        M4["'Performance' — unique"]
    end

    subgraph "Step 4: Consensus"
        C1["1. Onboarding — both ranked it #1"]
        C2["2. Dark mode — Alice #2, Bob #3"]
        C3["3. Performance — Bob #2 only"]
        C4["4. Mobile — Alice #3 only"]
    end

    A1 --> E
    B1 --> E
    E --> M1
    E --> M2
    E --> M3
    E --> M4
    M1 --> C1
    M2 --> C2
    M4 --> C3
    M3 --> C4

Notice how "Improve onboarding" and "Better onboarding" are recognized as the same idea despite different wording. The final ranking reflects both position importance (onboarding was #1 for both) and breadth of support.

Semantic Clustering

Traditional voting systems treat "improve UX" and "better user experience" as completely different ideas — splitting votes and distorting results. This system understands meaning, not just words:

flowchart TD
    subgraph "Different Words, Same Idea"
        A1["Improve onboarding"]
        A2["Better first-time experience"]
        A3["Streamline new user signup"]
    end

    A1 --> C["Grouped as ONE idea"]
    A2 --> C
    A3 --> C

    style C fill:#90EE90

This means stakeholders don't need to coordinate their vocabulary beforehand. Everyone can express priorities naturally, and the system finds the common ground.

Rank-Biased Weighting

Not all rankings are equal. If something is everyone's #1 priority, it should outrank something that only appears at #5. The system applies exponential weighting so top positions carry more influence:

flowchart TD
    subgraph "How Position Affects Weight"
        R1["#1 Priority → High Influence"]
        R2["#2 Priority → Medium Influence"]
        R3["#3 Priority → Lower Influence"]
        R4["...and so on"]
    end

    style R1 fill:#FF6B6B
    style R2 fill:#FFB347
    style R3 fill:#FFE066

This produces fair consensus: widely-shared top priorities rise to the top, while niche concerns from one stakeholder don't dominate.

Before and After

The transformation from scattered input to actionable consensus:

flowchart LR
    subgraph Before["Before: Scattered Input"]
        direction TB
        B1["Team A's list"]
        B2["Team B's list"]
        B3["Team C's list"]
        B4["Different words"]
        B5["Different order"]
        B6["Overlap unclear"]
    end

    T["Semantic<br/>RBO"]

    subgraph After["After: Clear Consensus"]
        direction TB
        A1["1. Top priority — everyone"]
        A2["2. Second priority — most"]
        A3["3. Third priority — some"]
        A4["Agreement scores"]
        A5["Support counts"]
    end

    Before --> T --> After

    style Before fill:#FFE4E1
    style After fill:#E1FFE4

Understanding Agreement

Beyond the consensus list, the system reveals how stakeholders relate to each other:

flowchart TD
    subgraph "Pairwise Agreement"
        P1["Alice ↔ Bob: 78% similar"]
        P2["Alice ↔ Carol: 45% similar"]
        P3["Bob ↔ Carol: 52% similar"]
    end

    subgraph "Alignment Analysis"
        Central["Alice & Bob: Central<br/>— close to consensus"]
        Outlier["Carol: Outlier<br/>— different priorities"]
    end

    P1 --> Central
    P2 --> Outlier
    P3 --> Outlier

    style Central fill:#90EE90
    style Outlier fill:#FFB347

This helps identify which stakeholders are aligned, which have unique perspectives worth discussing, and whether the group has broad agreement or significant divisions.

For a complete guide to reading and acting on results, see Interpreting Results.

Use Cases

• Requirements gathering from multiple stakeholders
• Feature prioritization across teams
• Strategic planning synthesis
• Survey response aggregation
• Collaborative decision making

License

MIT