Abstract
The retrieval-ranking paradigm has long dominated e-commerce search, but its reliance on query-item matching fundamentally misaligns with multi-stage cognitive decision processes of platform users. This misalignment introduces critical limitations: semantic gaps in complex queries, high decision costs due to cross-platform information foraging, and the absence of professional shopping guidance. To address these issues, we propose a Multi-Agent Cognitive Decision Framework (MACDF), which shifts the paradigm from passive retrieval to proactive decision support. Extensive offline evaluations demonstrate MACDF's significant improvements in recommendation accuracy and user satisfaction, particularly for complex queries involving negation, multi-constraint, or reasoning demands. Online A/B testing on JD search platform confirms its practical efficacy. This work highlights the transformative potential of multi-agent cognitive systems in redefining e-commerce search.
Introduction
Traditional e-commerce search engines operate on a "retrieval-ranking" basis: matching user keywords to product metadata and ranking results by predicted click-through rates. While effective for simple lookups, this approach fails to support the complex cognitive journey of a modern shopper. Users often face "semantic gaps" where intent isn't captured by keywords, high "decision costs" from manually comparing fragmented data, and a total lack of expert guidance for nuanced technical questions.
To bridge these gaps, we introduce the Multi-Agent Cognitive Decision Framework (MACDF). Rather than acting as a passive index, MACDF functions as a proactive shopping consultant, leveraging specialized agents to plan, research, and reason through user needs.
The MACDF Architecture
MACDF is built around a shared Memory System that maintains context and task states across a suite of specialized agents.
The framework employs several key components:
- Leader Agent: Orchestrates the process, delegating tasks and using a ReAct-inspired reflection mechanism to ensure quality.
- Guider Agent: Interacts with users to clarify ambiguous intents using Q-learning-based rewards.
- Planner Agent: Decomposes complex queries into a Directed Acyclic Graph (DAG) of manageable subtasks.
- ProductSearch & WebSearch Agents: Execute internal catalog searches and external web research (e.g., tech blogs, forums) to provide a holistic view of products.
- Decider Agent: Synthesizes all gathered information using multi-criteria decision-making theories to generate personalized recommendations with clear justifications.
Experimental Evaluation
We evaluated MACDF using the E-Commerce Cognitive Decision Benchmark (ECCD-Bench), a new dataset of 10,000 real-world JD.com queries spanning simple, complex, and consultative intents.
Offline Results
MACDF significantly outperformed both traditional online baselines and standard Retrieval-Augmented Generation (RAG) methods. The framework showed exceptional strength in handling logical constraints that typically baffle keyword-based systems.
As shown in the table above, for Negation Intent (e.g., "not a specific brand"), Accuracy@top5 soared from a baseline of 0.07 to 0.86. Similarly, Multi-Constraint Queries saw accuracy double from 0.41 to 0.80.
Online Impact
A two-week A/B test on JD.com involving 1% of live traffic confirmed the framework's real-world value:
- User Conversion Rate (UCVR): Increased by +6.5%, indicating more successful purchase decisions.
- User Click-Through Rate (UCTR): Improved by +3.9%.
- Average Reformulation Count (ARC): Decreased by -7.7%, suggesting the system understands user intent more accurately and earlier in the session.
Discussion and Future Work
While MACDF represents a significant leap in search intelligence, it introduces a "latency tax." Generating a full response currently takes approximately 15 seconds—far slower than traditional sub-second retrieval. Although streaming output mitigates this for the user (1.8s time-to-first-token), reducing computational overhead remains a primary focus. Future work will explore model compression, intelligent caching, and further ablation studies to optimize the multi-agent collaboration.
Summary updated