Exploring how ethics-neutral data processing decisions are not ethics-neutral by applying different document chunking strategies to data available through the UK Parliament API

• What We're Building (Non-Technical) - Understanding the project without code: why chunking decisions matter, how they shape political discourse, and what makes this uncomfortable
• Technical Implementation Plan - Complete technical architecture: stack decisions, chunking strategies, Neo4j schema, RAG pipeline design, and phase-by-phase implementation guide

• Roadmap Overview - Current status across all modules, next milestones, and recent wins
• Chunking Strategy Roadmap - Progress on the four chunking pipelines (2/4 complete), MVP milestones, and future features
• Neo4j Vector Storage Roadmap - Vector database implementation (complete), hybrid indexing strategy, and storage architecture
• User Interface Roadmap - SvelteKit frontend plans, comparative visualization, and citation display
• Parliament API Roadmap - API client implementation (complete), data ingestion, and testing
• Agents & Automation Roadmap - Slash commands and documentation workflows (complete)

• Parliament API Developer Guide - Complete reference for all UK Parliament API endpoints: Hansard, Members, Votes, Committees, and Bills
• API Testing Guide - Comprehensive testing workflows using Bruno and HTTPie, validation patterns, and integration strategies

• Roadmaps, Docs & Slash Commands Guide - Complete guide to the three-part documentation system: how roadmaps track progress, how docs explain concepts, and how slash commands automate updates

• Node.js 18+
• OpenAI API key (for text-embedding-3-large embeddings)
• Neo4j 5.11+ (for vector storage, required in later phases)

1. Clone the repository and install dependencies:

npm install

2. Copy the environment template and configure your API keys:

cp .env.example .env

3. Edit .env and add your OpenAI API key:

OPENAI_API_KEY=your_actual_openai_key_here

Test a single semantic chunking strategy (1024 tokens):

npm run test:chunking

Compare both semantic strategies side-by-side (1024 vs 256 tokens):

npm run test:chunking:compare

This will show:

• Chunk count differences between strategies
• Token granularity analysis
• Processing time comparisons
• Speaker and party distribution
• Sample chunks from each strategy

The Neo4j vector database stores chunks from all 4 chunking strategies using a hybrid indexing approach:

• Dual-label system: Each chunk has :Chunk + strategy label (:Semantic1024, :Semantic256, :Late1024, :Late256)
• 5 vector indexes: 4 strategy-specific + 1 unified for cross-strategy similarity
• Different embeddings per strategy: Semantic uses standard embeddings, late chunking uses blended embeddings (70% chunk + 30% debate context)

Step 1: Set up Neo4j Aura instance (see Neo4j Setup Guide)

Step 2: Verify connection and initialize schema (creates 5 vector indexes):

npm run test:neo4j:setup

Step 3: Populate with all 4 chunking strategies (~154 chunks from test data):

npm run test:neo4j:populate

Step 4: Test comparative vector search (queries all 4 strategies simultaneously):

npm run test:neo4j:search

This demonstrates the core experiment: how different chunking strategies retrieve different results for the same query. Test queries show 8-25% overlap between strategies, proving that chunking choices significantly impact retrieval.

The SvelteKit frontend provides an interactive comparative search interface:

Start the development server:

npm run dev

Access the UI: Open http://localhost:5173

Features:

• Comparative search: Query all 4 chunking strategies simultaneously
• Configurable results: Choose how many chunks (n) to retrieve per strategy (1-20)
• Divergence analysis: See mathematical breakdown of overlapping vs unique results
• Collapsible strategy views: Expand/collapse each strategy's results independently
• Rich citations: Party-colored speaker badges, Hansard references, similarity scores with tooltips
• Visual frames: Clear 2×2 grid showing "Early Chunking 1024/256" and "Late Chunking 1024/256"

Example queries:

• "What is the government's position on NHS funding?"
• "How did the opposition respond to the Prime Minister?"
• "What are the concerns about education policy?"

Anthropic