Idea
System improving ambiguous query intent understanding for multi-category marketplaces to boost search accuracy and relevance.
Research Paper
Core Innovation
This paper introduces an Agentic Multi-Source Grounded system that grounds LLM inference in a staged catalog entity retrieval pipeline and autonomous web search for cold-start queries. It outputs ordered multi-intent sets resolved by a configurable disambiguation layer, enabling extensibility and generalization across domains without modifying core architecture.
Why It Matters
Multi-category marketplaces face challenges in accurately interpreting short, ambiguous user queries, leading to poor search relevance and user experience. This system improves intent resolution by combining proprietary catalog data and real-time web knowledge, increasing accuracy and reducing errors. Its scalable, extensible design supports diverse marketplaces, enhancing search quality and customer satisfaction at scale.
Market Size (TAM)
$10–20B TAM for AI-powered search and intent understanding in multi-category marketplaces; $2–5B SAM from e-commerce, food delivery, and retail platforms. Driven by increasing demand for personalized, accurate search and growth of multi-vertical marketplaces.
Potential Customers & Pain Points
- Multi-category marketplaces – Difficulty resolving ambiguous user queries
- E-commerce platforms – Poor search relevance for sparse queries
- Food delivery services – Misclassification of user intent reducing order accuracy
- Retail aggregators – Challenges in integrating diverse inventory data for search.
Business Model
SaaS platform offering API access to multi-source grounded query intent understanding with tiered pricing based on query volume and customization level. Additional revenue from integration services and personalized disambiguation rule development.
Competitive Landscape
- Google Cloud AI Search
- Microsoft Azure Cognitive Search
- Amazon Kendra
- Algolia
- Coveo
Implementation Challenges
- Integration complexity with diverse proprietary catalogs
- Maintaining real-time web search accuracy and relevance
- Balancing multi-intent outputs with business policy constraints
- Scaling disambiguation rules across different marketplace verticals
Validation Strategy
- Pilot deployment with select multi-category marketplaces to measure search accuracy improvements
- A/B testing against existing search systems to quantify uplift in user engagement and conversion
- Longitudinal monitoring of query intent resolution accuracy and system scalability
- Customer feedback loops to refine disambiguation policies and personalization features
Research Paper Overview
Agentic Multi-Source Grounding for Enhanced Query Intent Understanding: A DoorDash Case Study
Summary
This paper presents a system that improves mapping ambiguous user queries to business categories by grounding large language model inference in catalog retrieval and autonomous web search. It outputs ordered multi-intent sets resolved by configurable disambiguation rules, enhancing accuracy and generalizability across marketplaces. Deployed at DoorDash, it significantly outperforms baseline and legacy systems, handling long-tail queries with high accuracy and serving most daily search impressions.