Delivery Hero implemented a sophisticated product matching system to identify similar products across their own inventory and competitor offerings. They developed a three-stage approach combining lexical matching, semantic encoding using SBERT, and a retrieval-rerank architecture with transformer-based cross-encoders. The system efficiently processes large product catalogs while maintaining high accuracy through hard negative sampling and fine-tuning techniques.
# Delivery Hero's Product Matching System Using LLMs
Delivery Hero, a global online food delivery and quick commerce company, developed a sophisticated product matching system to address the challenges of identifying similar products across their inventory and competitor offerings. This case study details their journey in implementing a production-grade LLM-based solution for semantic product matching.
## Business Context and Problem Statement
- Delivery Hero needed to identify similar products between their inventory and competitor offerings
- The system needed to help develop pricing strategies and understand product variety differences
- Additional use case included identifying duplicate items within their own product range
- The solution focused on matching products using product titles, with potential for expansion to include images and other attributes
## Technical Implementation Journey
### Initial Approach: Lexical Matching
- Implemented basic text matching using bag-of-words approach
- Enhanced with Term Frequency (TF) and Inverse Document Frequency (IDF)
- Utilized BM25 similarity scoring
- Key advantages:
- Limitations:
### Advanced Implementation: Semantic Encoder
- Utilized SBERT (Sentence-BERT) library
- Based on pre-trained transformer LLM
- Customization through fine-tuning:
- Technical advantages:
- Implementation challenges:
### Production Architecture: Retrieval-Rerank System
- Implemented two-stage architecture for optimal performance
- Stage 1 - Retrieval:
- Stage 2 - Reranking:
## MLOps and Production Considerations
### Model Training and Optimization
- Implemented hard negative sampling strategy:
- Fine-tuning pipeline:
### System Architecture Considerations
- Built scalable pipeline handling large product catalogs
- Implemented efficient data processing workflows
- Balanced system resources between:
- Designed for production deployment with:
### Production Deployment Strategy
- Implemented staged rollout approach
- Created monitoring systems for:
- Established feedback loops for continuous improvement
- Developed maintenance protocols for:
## Technical Outcomes and Benefits
- Successfully identified similar products across different catalogs
- Improved pricing strategy development
- Enhanced product assortment management
- Achieved balance between accuracy and computational efficiency
- Created scalable solution for growing product catalog
## Future Enhancements and Scalability
- Potential expansion to include:
- Planned improvements:
## Key Learning Points
- Importance of balanced approach between speed and accuracy
- Value of multi-stage architecture in production systems
- Critical role of hard negative sampling in model improvement
- Necessity of efficient resource utilization in production
- Benefits of incremental implementation approach
## Technical Infrastructure
- Utilized SBERT for semantic encoding
- Implemented transformer-based cross-encoders
- Developed custom fine-tuning pipelines
- Created efficient data processing workflows
- Built scalable production architecture
This implementation demonstrates a sophisticated approach to deploying LLMs in a production environment, balancing technical requirements with business needs while maintaining system efficiency and scalability.
Start deploying reproducible AI workflows today
Enterprise-grade MLOps platform trusted by thousands of companies in production.