Semantic Car Search:A Vector-Based Approach

1.1 Motivation: The Automotive Semantic Gap

Automotive catalogs exhibit deep hierarchical taxonomies:

However, car spotters, collectors, and casual enthusiasts almost never formulate queries aligned with this strict canonical hierarchy. Real-world queries exhibit high entropy across four distinct failure modes for classical lexical systems:

• Slang & Nicknames

  Users enter colloquial terms like "Beamer" (BMW), "Merc" (Mercedes), "Rari" (Ferrari), "Miata" (Mazda MX-5), or "G-Wagon" (G-Class).

• Generational Variations

  A search for "911 Turbo" encompasses 50 years of divergent chassis codes (930, 964, 993, 996, 997, 991, 992) which share no lexical tokens with "Turbo".

• Typos & Phonetic Errors

  Mobile camera users typing on the move submit severe typos: "LAmborgini Huracan", "Ferari Testarosa", or "Porshe Cayan".

• Multilingual Descriptive Intent

  Non-English queries describing mechanical attributes without brand labels, e.g. "voiture de sport italienne v10" or "break allemand puissant".

1.2 Problem Statement & Production SLAs

To power Wheeloh's mobile vehicle identification experience, the search engine must operate within three strict engineering Service Level Agreements (SLAs):

2.1 Vector Embeddings & Document Serialization

Each vehicle entity in the Wheeloh database is serialized into a structured contextual representation prior to vectorization:

# Contextual document serialization for automotive entities
def serialize_vehicle(car: dict) -> str:
return (
f"Make: {car['brand']} | "
f"Model: {car['model']} | "
f"Generation: {car.get('generation', 'N/A')} | "
f"Version: {car['version']} | "
f"Powertrain: {car.get('powertrain', 'N/A')}"
)
# Unit-normalized dense embedding generation (1,536 dimensions)
doc_vector = model.encode(
input=serialize_vehicle(vehicle),
normalize_embeddings=True  # Enforces ||d||_2 = 1.0
)
# Output: ndarray(shape=(1536,), dtype=float32, norm=1.0)

The encoder projects these serialized strings into a continuous 1,536-dimensional metric space where automotive semantic relationships correspond to directional proximity on a hypersphere.

2.2 Similarity Computation & SIMD Acceleration

Semantic similarity between an incoming query vector q and candidate vehicle vector d is governed by cosine similarity:

Because all vehicle representations d and query vectors q are pre-normalized to unit L2 norm (||q|| = 1 and ||d|| = 1), the denominator simplifies to unity. The similarity score reduces strictly to a linear inner dot product:

Matrix multiplication across the entire dataset of 22,180 indexed vehicles completes in 1.6 milliseconds via AVX2 SIMD fused multiply-add (FMA) instructions. Top-k candidates are isolated using a linear-time partition:

# Vectorized SIMD dot-product and top-k retrieval in NumPy
import numpy as np
# D: Pre-loaded matrix of shape (22180, 1536), dtype float32 (34 MB in RAM)
scores = np.dot(D, q)  # 1.6 ms via AVX2 SIMD / BLAS
# Linear-time Top-K selection without full O(N log N) sorting
k = 5
candidate_indices = np.argpartition(scores, -k)[-k:]
top_k_sorted = candidate_indices[np.argsort(-scores[candidate_indices])]
# Result: Top-5 candidate models isolated in 1.8 ms total execution time

2.3 Latent Manifold Projection (PCA)

To inspect cluster topology and verify semantic coherence, we compute an orthogonal linear transformation via Principal Component Analysis (PCA). The 1,536-dimensional latent space is projected onto the top two eigenvectors that capture maximum spatial variance, yielding a 2D coordinate system: