Semantic Commerce Graph

Challenge

A major e-commerce retailer with 10M+ SKUs was losing customers to competitors with better product discovery. Their keyword-based search couldn’t understand natural language queries like “comfortable shoes for standing all day” or “gift for tech-savvy teenager.”

Solution

We implemented a Retrieval-Augmented Generation (RAG) architecture that combines:

Semantic product embeddings generated from product descriptions, reviews, and attributes
Hybrid search merging vector similarity with traditional filters (price, availability, brand)
Query understanding using LLMs to extract intent and expand queries
Personalization layer incorporating browsing history and purchase patterns

Technical Implementation

Search Architecture

User Query: "waterproof jacket for hiking in rain"
                    │
                    ▼
┌─────────────────────────────────────────────────────────┐
│              Query Understanding (Gemini-3)              │
│  Intent: outdoor_apparel                                 │
│  Attributes: waterproof, hiking, rain_protection         │
│  Expanded: "rain jacket", "hiking shell", "gore-tex"     │
└─────────────────────────────────────────────────────────┘
                    │
        ┌───────────┴───────────┐
        ▼                       ▼
┌───────────────┐       ┌───────────────┐
│ Vector Search │       │ Filter Engine │
│  (Pinecone)   │       │  (Postgres)   │
└───────────────┘       └───────────────┘
        │                       │
        └───────────┬───────────┘
                    ▼
┌─────────────────────────────────────────────────────────┐
│              Hybrid Ranking & Reranking                  │
│         (Cross-encoder + Business Rules)                 │
└─────────────────────────────────────────────────────────┘
                    │
                    ▼
            Top 24 Products

Embedding Strategy

Multi-modal embeddings: Text + image features from product photos
Review synthesis: Aggregated sentiment and feature mentions
Attribute injection: Structured data (size, color, material) as text

Results

After 90 days of A/B testing across 2M daily active users:

2.4x improvement in search-to-purchase conversion
31% reduction in “no results” queries
18% increase in average order value
47% reduction in search refinements needed

Customer Feedback

“The new search actually understands what I’m looking for. I searched for ‘laptop for video editing under $1500’ and got exactly what I needed on the first page.” — Beta tester feedback