Car Recommendation System

An AI-powered recommendation engine combining semantic search, vector databases, and natural language understanding to match users with the right vehicles

The Challenge

Static filter search: Most car searches rely on filters (make/model/specs) assuming users already know exactly what they want
Unclear intent: Many buyers start with only a rough idea or simply a set of needs
Outcome: Friction, drop-offs, and missed relevant alternatives

💡 The Core Problem:
Users think in use-cases. They start with why (use-cases), not what (specs). Marketplaces do the opposite.

The Solution

1

Input: Natural Language or Form

Users either describe their situation in natural language (e.g., "family needing space for kids and luggage") or fill a structured form with key features like budget, seats, fuel and usage.

→

2

Interpretation & Structuring

For natural language input, GPT interprets intent and asks clarifying questions; for form input, the structured features are used directly without additional AI interpretation.

→

3

Smart Matching

CRS returns matching vehicles + convincing alternatives via semantic retrieval

🎯 Result:
Users can choose between a conversational, guided search via natural language or a fast expert form. In both cases, preferences are transformed into structured data for precise matching instead of relying on rigid keyword filters.

How It Works

📝 Input → Structure

Users either describe their needs in natural language or fill a structured form. Both paths are transformed into a unified preferences JSON (seats, budget, fuel, gearbox, usage…)

🔍 Vector Search

Vector search in OpenSearch finds similar vehicles (kNN), even with incomplete inputs

⚙️ Hybrid Filtering

Hard filters + ranking + explanation produce results

🎯 Guided Experience

Users experience a competent, human-like consultation instead of tedious manual filtering

Approach Example: Text → Structured Query

👤 User Input

"Family of 5, city commuting"
"Need 7 seats, automatic"
"Budget ≤ €18k, max 120k km"

🤖 CRS Preference JSON

{
  "numberOfSeats": 7,
  "gearbox": "automatic",
  "price_max": 18000,
  "mileage_max": 120000,
  "usage": "city",
  "missing": ["fuel", "bodyType"]
}
                    

→ Follow-up: "SUV or van? Fuel preference?"

Technical Deep Dive: Prompt & Dialog Strategy

System prompt defines a strict, machine-readable JSON schema for CarPreferences.
Dialog policy: in each turn the assistant either asks one clarifying question or returns a complete preferences JSON – never a mixture.
Chat memory keeps track of which fields are already known and which are still missing.

Technical Deep Dive: Semantic Vector Search

Semantic Vector

Embedding space

Embeddings of car descriptions and a matching user query vector in the same space

→

OpenSearch

kNN vector search

→

Top-k

Ranked Cars

Semantic similarity scoring + metadata filters (price, mileage, year)

Single semantic space: Each car and the user query are embedded into the same vector space, so distance directly reflects relevance.
kNN-based retrieval: OpenSearch kNN finds the nearest neighbors for the query vector and combines them with standard filters on price, mileage and year.
Robust to vague input: Works even if users cannot specify exact technical specs and only describe their situation in natural language.

End-to-End Architecture (Numbered Flow)

1

Ingestion & Enrichment

🌐

Marketplaces

mobile.de / AutoScout24

📦

S3 Bucket

Raw Ads (JSON)

⏰

EventBridge

Daily Schedule

↓

⚡

Lambda

Orchestrator

🔧

AWS Glue ETL

Normalize + Enrich

🧬

Embedding Generator

Vector Creation

2

Storage & Indexing

💾

DynamoDB

Structured Features

🔍

OpenSearch

kNN Index (Vectors)

🗄️

S3

Raw Archive

3

Query & User Experience

💬

Streamlit UI

Form + Chat Mode

🤖

GPT-4o

Intent → JSON

⚙️

Query Engine

Vector + Filters

🚀 Key Architecture Features:

AWS-managed, largely serverless architecture
Daily automated data refresh pipeline
Optimized for low-latency retrieval (vector similarity + metadata filters)
Modular components (easy to swap GPT/embeddings/index providers)

Evaluation: Test Design & Data Basis

62k

Total Dataset
Vehicle Listings

1,000

Test Set
Randomly Selected

3

Test Scenarios
Feature Combinations

🧪 Reproducible Test Procedure

For each test vehicle, exactly one search query is generated
Search queries based exclusively on technical attributes (no make/model)
Independent search for each of 1,000 vehicles, defined by objective characteristics

What is Measured?

📊 Hit Accuracy: How often does CRS find the correct vehicle based only on features?
📍 Ranking Position: At which position does the vehicle appear? (Rank 1, Top 3, Top 5, Top 10)
🔧 Robustness: How stable is accuracy when features are missing or incomplete?

Evaluation: Definition of Test Cases

Test 1 – 9 Features

Complete data scenario

BodyType NumberOfDoors FirstRegistration GearBox NumberOfSeats Fuel Power DriveType CubicCapacity

Test 2 – 6 Features

Realistic partial data

BodyType FirstRegistration GearBox NumberOfSeats Fuel Power

Test 3 – 3 Features

Sparse data scenario

BodyType GearBox Fuel

📌 Design Rationale: The tests simulate realistic listing completeness across three levels: core-feature complete, partially complete, and sparse listings (within the selected feature subset).

Evaluation: Results Overview

Test Scenarios	Rank 1 (Hit@1)	Top 3 (Hit@3)	Top 5 (Hit@5)	Top 10 (Hit@10)
Test 1 – 9 features	71.4%	90.7%	94.5%	97.9%
Test 2 – 6 features	59.3%	83.6%	89.7%	95.0%
Test 3 – 3 features	19.1%	39.1%	51.3%	68.4%

📊 Interpretation: For each test scenario, the table shows in what percentage of the 1,000 searches the target vehicle appears on Rank 1, within the Top 3, Top 5, or Top 10 of the result list. Each row corresponds to 1,000 searches for the respective test vehicles.

Evaluation: Hit Rates by Rank

Hit@1 (Rank 1)

Test 1 (9 features)

71.4%

Test 2 (6 features)

59.3%

Test 3 (3 features)

19.1%

Hit@10 (Top 10)

Test 1 (9 features)

97.9%

Test 2 (6 features)

95.0%

Test 3 (3 features)

68.4%

Evaluation: Top 10 Hit Rate Analysis

979

Test 1: Vehicles in Top 10
(out of 1,000)

950

Test 2: Vehicles in Top 10
(out of 1,000)

684

Test 3: Vehicles in Top 10
(out of 1,000)

Performance Characteristics

Stable performance: Hit@10 remains high, dropping only from 97.9% to 95.0% when reducing features from 9 to 6 (only 2.9 percentage points)
Robustness with sparse data: Even with only 3 features (BodyType, GearBox, Fuel), the system finds 68.4% of vehicles in Top 10
Practical implication: System works well even with incomplete marketplace listings

🎯 Core Metric: Hit@10 = 97.9%
In the most complete scenario (9 features), nearly all vehicles (979/1000) appear within the first 10 results, demonstrating excellent retrieval accuracy.

Evaluation: Technical Interpretation

🔍 Vector Search Performance

The system reliably re-identifies vehicles using vector search and similarity metrics, making it much more flexible than purely static filter logic.

⚙️ End-to-End Validation

Evaluation covers the entire technical chain: DynamoDB storage → OpenSearch kNN indexing → result quality metrics.

Key Technical Insights

Semantic vector search works: A single embedding vector per car delivers reliable re-identification and ranking of relevant vehicles.
kNN search performs well: Vector search in OpenSearch achieves high recall without requiring full exhaustive scans.
Metadata filtering effective: Price, mileage, and year constraints work seamlessly together with vector search.
Reproducible results: Tests show that the end-to-end pipeline (DynamoDB → OpenSearch kNN → ranking) is correctly integrated and stable.

💡 Practical Takeaway: The dual vector strategy enables the system to handle real-world scenarios where data is incomplete, while still delivering highly relevant results. This is a significant advantage over traditional keyword-based or pure filter systems.

Impact & Business Value

📈 User Experience

More relevant recommendations
Reduced friction in search
Guided consultation experience

💰 Business Metrics

Increased conversion likelihood
Higher user engagement
Reduced drop-off rates

🚀 Foundation for Scale

White-label integration friendly: Can be integrated into existing marketplaces via API or embeddable widget
Multi-domain potential: Architecture extends beyond cars (real estate, e-commerce, job matching)
Production-ready: AWS-managed infrastructure designed for reliability and scale

🎯 Core Achievement

CRS demonstrates how intelligent data architecture and generative AI can transform traditional filter-based search into a precise, user-centered recommendation experience.

Next: White-Label + Multi-Domain Growth

1

White-Label Integration

API + embeddable widget for seamless marketplace integration

→

2

Partner Rollout

KPI optimization: conversion rate, CTR, drop-off reduction

→

3

Multi-Domain Expansion

Extend to real estate, jobs, e-commerce via configurable schema

🔧 Technical Extensibility

Modular architecture: Easy to swap LLM providers (GPT → Claude/Llama), embedding models, or vector databases
Configurable schema: JSON-based preference definitions adapt to different product domains
A/B testing ready: Built-in analytics hooks for measuring recommendation quality improvements