Semantic Product Recommender

Project Overview

This semantic product recommender helps a leading US procurement platform provide relevant alternative products when items are out of stock or overpriced. The system transforms the product discovery experience through advanced semantic understanding and metadata-aware search techniques.

Technical Solution

System Architecture

We designed a sophisticated recommendation engine with:

1. Semantic Core Development:
   • Advanced embedding techniques for product similarity
   • Hybrid search combining semantic and metadata filters
   • Optimized context window size for embedding model effectiveness
   • Product text restructuring to prioritize distinctive features
2. Technical Optimization:
   • Multiple text pooling methods with first-token pooling proving superior
   • Expanded candidate pools for recommendation diversity
   • Metadata-based filtering for price ranges and manufacturer data
   • Benchmark testing framework for continuous improvement
3. Enhanced Search Features:
   • Price sensitivity awareness
   • Manufacturer relationship mapping
   • Category-aware recommendations
   • Usage pattern matching

Optimization Process

The solution incorporates several key innovations:

• Context window optimization increasing from default to 32,768 tokens
• Text pooling experimentation (mean, first-token, last-token)
• First-token pooling providing 67% improvement in relevance
• Product text restructuring to place distinctive features first
• Hybrid search combining text similarity with attribute matching

Implementation Challenges

Key challenges we addressed included:

• Balancing semantic similarity with critical product attributes
• Maintaining reasonable query response times
• Creating a robust benchmark dataset for evaluation
• Handling diverse product categories beyond initial test cases
• Optimizing text representation without expensive model fine-tuning
• Dealing with incomplete and inconsistent product metadata

Business Impact

The system delivered substantial value across multiple dimensions:

Recommendation Quality Metrics

• Mean Reciprocal Rank (MRR): Improved from 0.3000 to 0.7000 (133% increase)
• Recommendation Position: Relevant products now appear in positions 1-2 vs. previous 3-4
• Mean Average Precision (MAP@10): Increased from 0.36 to 0.65
• Query Response Time: Maintained reasonable 1.39 seconds despite more sophisticated processing
• Recommendation Diversity: 46% improvement in product category variation
• Relevance Score: 87% in blind expert evaluations (vs. 52% for previous system)

User Experience Metrics

• Search Time Reduction: 73% less time spent searching for alternative products
• Successful Substitutions: 28% increase when preferred items unavailable
• Productivity Improvement: Procurement specialists now manage 35% more purchase orders
• User Adoption: 68% of users regularly utilize recommendation features within 90 days

Business Outcome Metrics

• Purchase Order Efficiency: 35% more POs processed with same resources
• Cost Savings: 26% average savings when selecting recommended alternatives
• Process Exception Reduction: 37% fewer procurement process exceptions
• Abandoned Purchase Reduction: 12% decrease in abandoned procurement journeys