Idea
Search system expanding e-commerce item discovery to boost user satisfaction and marketplace exposure.
Research Paper
Core Innovation
This paper introduces a two-stage hybrid architecture combining large closed-weight LLMs and a finetuned small language model trained via LoRA adapters and teacher-student distillation. This approach balances the cost-quality tradeoff in generative retrieval, enabling scalable intent-conditioned recall expansion that improves both semantic quality and downstream retrieval effectiveness.
Why It Matters
E-commerce platforms struggle to surface substitute and complementary items, limiting user satisfaction and sales. This system improves item discoverability by broadening search recall with related intents, increasing purchase opportunities and balancing marketplace supply exposure. It scales efficiently to cover most query traffic while controlling inference costs, making it practical for large marketplaces.
Market Size (TAM)
$20–50B TAM for e-commerce search and recommendation; $5–10B SAM from large online marketplaces and grocery retailers. Driven by growing demand for personalized search and increased competition in online retail.
Potential Customers & Pain Points
- E-commerce marketplaces – Limited item discoverability reduces sales
- Grocery retailers – Need to surface substitutes and complements
- Online search platforms – High inference cost limits recall expansion
- Long-tail sellers – Lack of exposure in search results
Business Model
SaaS platform or API licensing to e-commerce marketplaces and retailers, with tiered pricing based on query volume and model usage. Potential for revenue share on incremental sales driven by improved discoverability.
Competitive Landscape
- Amazon Search
- Google Shopping
- Coveo
- Algolia
- Bloomreach
Implementation Challenges
- Integration complexity with existing search infrastructure
- Balancing recall expansion without degrading precision
- Computational cost and latency constraints for real-time search
- Adoption resistance due to changes in ranking and exposure dynamics
Validation Strategy
- Pilot deployment with select e-commerce partners to measure session-level purchase lift
- A/B testing comparing baseline search with discovery-augmented search
- Human evaluation of generated intents for semantic relevance
- Cost-benefit analysis of inference cost versus revenue impact
Research Paper Overview
Improving Item Discoverability in e-Commerce Search via Related Intent Generation
Summary
This paper presents a scalable discovery-augmented search system for e-commerce that expands candidate recall by generating implicit user intents, improving item discoverability beyond exact query matches. It uses a two-stage hybrid model combining large closed-weight LLMs for head queries and a finetuned small language model for tail queries, balancing cost and quality. Evaluations show increased discovery coverage from 60% to 80% of query traffic at 30% of the teacher model's cost, enhancing both search relevance and marketplace exposure for long-tail supply.