Phong Trang

AI Semantic Search for New Car Sales

By Phong Trang - Software Product Manager - thanhphong.trang@gmail.com

Overview

The Challenge: Transform Volvo’s car discovery experience by building an intelligent search system that understands customer intent, bridging the gap between natural language queries and technical car specifications.

The Impact: Achieved 85% search accuracy (NDCG score), increasing 19% to Car Listing Pages, and make search user 11% more likely to complete a car purchase.

Technical Stack:

OpenAI Python Cohere Elasticsearch GraphQL Vector Search Google Analytics LLM

Key Metrics:

Semantic Search Architecture

Problem Statement

💔 What Was Breaking

🚙 Complex Purchase Decisions

Cars represent the second-largest purchase after homes, with hundreds of configurations and features. Customer research typically spans weeks, but our existing search made it difficult to find the right match in inventory.

🏗️ Legacy System Limitations

The backend GraphQL system provided comprehensive car data but lacked semantic search capabilities:

🗣️ Semantic Language Gap

A critical disconnect existed between how customers search and how we catalog cars:

🛒 Direct-to-Consumer Strategy Risk

The UK market pilot for transparent, haggle-free car sales required self-service discovery capabilities that our legacy search couldn’t support, threatening the strategic initiative.

💡 The Opportunity

  • Revenue Recovery: Capture lost sales from customers who couldn’t find suitable cars in our existing inventory through better discovery
  • Competitive Advantage: Differentiate Volvo’s direct-to-consumer experience with Google-like search intelligence while competitors rely on traditional dealership browsing
  • Inventory Optimization: Convert hidden quick-delivery stock into competitive advantage by making 3-week delivery cars easily discoverable

🕵️ Discovery & Research

Data-Insight-Belief-Bet Framework

📐 Data (what we can reliably measure)

The baseline facts:

💠 Canonical metrics & formulas

👥 User Research Insights

🗺️ Customer Journey Mapping

Identified three primary search intents:

  1. 🛒 Purchase Intent: New car buyers with specific needs
  2. 📚 Information Seeking: Research-phase customers exploring options
  3. 🔧 Existing Owner Support: Current customers seeking information about their vehicles

🔬 Technical Insight

🧠 Semantic Router Investigation

Researched industry-standard approaches, focusing on the semantic router methodology:

  1. Intent classification through utterance databases
  2. Vector similarity matching using embedding models
  3. Real-time query processing and ranking

⚖️ Embedding Model Evaluation

Used NDCG (Normalized Discounted Cumulative Gain) as the primary evaluation metric to measure ranking quality across 1,200+ test queries to benchmark:

💡 Solution Strategy

🎯 Strategic Approach

🧭 Intent Recognition First

Rather than treating this as a pure search problem, I positioned it as an intent recognition challenge. The key insight: understand what customers want, not just what they type.

🏗️ Semantic Router Architecture

Designed a three-layer system:

  1. 🎭 Intent Classification: Categorize user queries into predefined intents (EV interest, car type preferences, feature requirements)
  2. 🔗 Semantic Matching: Use vector similarity to find relevant cars in inventory
  3. 🎯 Intelligent Reranking: Apply cross-encoder models for final result optimization

🚀 Key Product Decisions

🔀 Multi-Modal Search Approach

🌍 Multilingual Strategy

For German market expansion:

⚡ Performance-First Design

🧪 Interactive Prototype

Built a real-time demo application using semi-live UK market data to showcase the semantic search capabilities and validate performance with stakeholders.

Semantic Search Demo - Lexical vs Semantic

Comparison:

💡 The Power of Context

Notice how semantic search correctly identifies that “family car” means larger SUV models (XC90) rather than compact cars (EX30), and “sun roof” maps to “panoramic roof” in the vehicle specifications - exactly the kind of intelligent interpretation that traditional search systems miss.

🛠️ Implementation

System Architecture

Implementation Architecture

Final Product

🪙 Cost-Effective Implementation

📊 Results & Impact

🎯 Performance Metrics

🏆 Search Accuracy Achievement

🥇 Model Performance Ranking

Model Performance Comparison

🌍 Cross-Market Validation

💼 Business Impact

🎯 Conversion Funnel Improvements

😊 Customer Experience Improvements

⚡ Technical Achievements

📈 Operational Efficiency

Tech Stack

Skills Demonstrated