Back to Blog
Technical

PawsomeChoice Technical Architecture

A deep-dive into the architecture behind PawsomeChoice: vector embeddings with Pinecone, RAG-powered article generation, AWS Bedrock, and C4 model diagrams.

Haris KabiljagicJanuary 9, 202610 min read

In Part 1 of this series, I shared the entrepreneurial journey of building PawsomeChoice - a dog product comparison platform I built solo with AI assistance.

This post goes deeper into the technical architecture: how semantic search actually works, the RAG pipeline that generates articles, and the infrastructure choices that let one person maintain 40,000+ products.

Haven't read Part 1? Start with How I Built a Price Comparison Platform Solo for the full entrepreneurial story.

System Architecture Overview

The architecture follows the C4 model - a hierarchical approach that lets you zoom in from high-level context down to individual components. The key principle: automation-first, with minimal manual intervention required.

System Context (C4 Level 1)

The high-level view showing how users interact with PawsomeChoice and its integration with stores and affiliate networks.

Tech Stack Decisions

Every technology choice was made with solo maintainability in mind. Here's what powers PawsomeChoice:

LayerTechnologyWhy This Choice
FrontendNext.js + ReactSSR for SEO, hosted on AWS Amplify
BackendECS + LambdaContainerized services and serverless functions
OrchestrationStep FunctionsComplex workflows and data pipelines
EventsEventBridgeScheduled tasks and event-driven automation
SearchMeiliSearchFast, typo-tolerant, self-hosted on AWS
Vector DBPineconeManaged vector search, scales well
AI/MLAWS BedrockMultiple models (Claude, Titan, etc.) pay-per-use

Container Architecture (C4 Level 2)

Detailed view of the system components including the UI, backend services, search engine, and database layers.

The Heart: AI-Powered Semantic Search

Traditional search matches keywords. PawsomeChoice understands meaning.

How It Works

  1. Every product is vectorized using AI embeddings
  2. Vectors are stored in Pinecone (purpose-built vector database)
  3. User query → converted to vector → similarity search
  4. Results ranked by semantic relevance, not keyword matching

Example

Search: "best dry food for puppies with sensitive stomachs"

Traditional search: Looks for exact words "dry" + "food" + "puppies" + "sensitive"

Semantic search: Understands intent → finds "grain-free puppy food for digestive issues" even without exact keyword match

What are vector embeddings? Mathematical representations of text that capture meaning. Similar concepts have similar vectors - allowing "grain-free puppy food" to match "food for young dogs with allergies".

The Article Generation Pipeline

PawsomeChoice automatically generates helpful articles using a sophisticated RAG (Retrieval-Augmented Generation) pipeline. Here's how it works:

Step 1: Topic Selection

Identify content gaps, analyze search queries without good results, queue topics for generation.

Step 2: Product Discovery (RAG)

Query vectorized product database for semantically relevant products. Initial pool: 20-50 candidates.

Step 3: Hybrid Scoring Algorithm

Products are ranked using a two-tier hybrid scoring system:

Final Score
Semantic Match 60%
Quality 40%
Quality Score Breakdown
Rating
50%
Review count
30%
Store coverage
10%
Rating sources
10%

Step 4: Content Generation (AWS Bedrock)

Structured prompt with product data → generate intro, quick facts, FAQ, product breakdowns. Tailored for Swedish audience, SEO + AEO optimized.

Step 5: Publishing & Distribution

Auto-publish to PawsomeChoice, distribute to social (Instagram, Pinterest). AI-generated header images.

Key insight: These aren't "AI slop" articles. Each references actual products on the platform, includes real pricing, and provides genuine value. The AI augments human judgment - it doesn't replace it.

Component Diagram: Article Service (C4 Level 3)

A deep dive into the AI-powered Article Recommendation & Generation Service - the heart of PawsomeChoice's intelligent content system using vector embeddings, RAG, and AWS Bedrock.

Product Pages: Data Transparency

Every product page on PawsomeChoice is designed for trust and transparency:

  • Descriptions cleaned and normalized from messy affiliate feeds
  • Specifications extracted and structured for easy scanning
  • Price history charts to spot real deals vs fake discounts
  • Related products via vector similarity (not just category)
  • Rating aggregation with source breakdown

"Everything is transparent. No hidden algorithms boosting certain products. No paid placements. Just data, clearly presented."

Scaling Considerations

While currently focused on Sweden with 16 stores, the architecture supports international expansion. Adding a new country means configuring new affiliate feeds and translating the UI - the core platform handles everything else.

The niche focus on dog products is intentional: rather than being a mediocre general comparison site, PawsomeChoice aims to be the best resource for dog owners specifically.

Missed the Entrepreneurial Story?

Read Part 1 to learn how this project started as a frustrated dog owner's side project and the lessons learned building it while keeping a full-time job.

Read Part 1: The Entrepreneur Story

See It in Action

Try the semantic search yourself. Search for something natural like "toys for aggressive chewers" and watch the AI understand what you need.

About the Author: Haris Kabiljagic is a principal engineer and cloud architect with 20+ years of experience. He built PawsomeChoice as a solo side project to explore practical AI applications in production systems. Connect on LinkedIn to discuss architecture and AI-assisted development.