💻 computer science

CogSearch: A Cognitive-Aligned Multi-Agent Framework for Proactive Decision Support in E-Commerce Search

This paper introduces CogSearch, a cognitive-aligned multi-agent framework that transforms e-commerce search from passive retrieval into proactive decision support, demonstrating significant improvements in conversion rates and reduced decision costs through extensive online testing on JD.com.

Original authors: Zhouwei Zhai, Mengxiang Chen, Haoyun Xia, Jin Li, Renquan Zhou, Min Yang

Published 2026-03-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zhouwei Zhai, Mengxiang Chen, Haoyun Xia, Jin Li, Renquan Zhou, Min Yang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are walking into a massive, chaotic warehouse (like a giant e-commerce site) looking for a specific item. In the old days, you would shout a keyword like "camera," and a robot would instantly hand you a stack of 1,000 boxes. You'd have to open every single one, read the labels, compare prices, and figure out which one fits your needs. This is exhausting, confusing, and often leads to you giving up or buying the wrong thing.

CogSearch is like hiring a team of four expert personal shoppers who don't just fetch boxes; they actually think with you to help you make the best decision.

Here is how this "Cognitive-Aligned Multi-Agent Framework" works, broken down into simple parts:

1. The Problem: The "Passive" Search Engine

Traditional search engines are like vending machines. You press a button (type a query), and they give you a list of items. They don't care if you are confused, if you have a tight budget, or if you need a camera for hiking vs. a camera for vlogging. They just show you what matches the words. This puts all the mental work (cognitive friction) on you, the user.

2. The Solution: The "CogSearch" Team

CogSearch replaces the vending machine with a collaborative team of four specialists who work together to guide you from "I need something" to "I bought the perfect thing."

🧠 The Planner (The Architect)

What it does: When you say something vague like "I need a gift for my dad," the Planner doesn't just search for "gift." It breaks your request down into a to-do list.
The Analogy: Think of the Planner as a project manager. If you say "I want to build a treehouse," the manager doesn't just hand you wood. They create a blueprint: Step 1: Find the right wood. Step 2: Check the weather. Step 3: Find a hammer.
In CogSearch: It turns your messy thoughts into a structured plan, deciding what information it needs to gather next.

🕵️ The Executor (The Detective)

What it does: This agent goes out and gathers all the facts. It doesn't just look inside the store; it checks the internet, reads reviews, and even checks external tools (like weather forecasts or logistics).
The Analogy: Imagine a detective who doesn't just look at the crime scene but interviews neighbors, checks the news, and looks at satellite maps.
In CogSearch: It fetches product lists from the store, reads expert reviews online, and checks if the item is in stock or if shipping will be delayed due to rain.

🧭 The Guider (The Coach)

What it does: Once the Detective has a pile of information, the Coach helps you narrow it down. It asks smart questions and creates custom filters based on what you actually care about.
The Analogy: Think of a personal trainer. If you say "I want to get fit," they don't just hand you a generic workout sheet. They ask, "Do you have knee pain? Do you have 10 minutes or an hour?" and then build a custom plan.
In CogSearch: Instead of showing you 1,000 cameras, it says, "Since you mentioned you go hiking, let's filter for cameras that are lightweight and waterproof." It proactively suggests, "Did you know you also need a tripod for this?"

⚖️ The Decider (The Judge)

What it does: This is the final boss. It takes all the gathered info, your budget, your history, and the Coach's advice, and makes a final recommendation with a clear explanation.
The Analogy: This is like a wise judge in a courtroom. They listen to all the evidence (the Detective's facts), consider the laws (your constraints), and deliver a verdict: "Based on your budget and needs, this specific camera is the winner because..."
In CogSearch: It doesn't just give you a ranked list. It says, "I recommend Camera X. It's slightly more expensive, but the battery lasts 2 hours longer, which is perfect for your hiking trip."

3. The Results: Why It Matters

The paper tested this system on JD.com (one of China's biggest online stores). The results were impressive:

Less Stress: Users spent 5% less time making decisions. They didn't have to click through as many pages or search as many times.
More Sales: Overall, more people bought things (a 0.41% increase in conversion).
The "Hard" Stuff: For difficult decisions (like buying a gift or comparing complex electronics), sales jumped by 30%.

The Big Picture

CogSearch changes the goal of search.

Old Way: "Here is a list of things that match your words. Good luck figuring it out."
CogSearch Way: "We understand what you really need. We've done the research, checked the reviews, and here is the best choice for your specific situation, along with the reasons why."

It turns the search engine from a library catalog (where you have to find the books yourself) into a knowledgeable librarian who walks you to the exact shelf, pulls out the book, and explains why it's the one you should read.

1. Problem Statement

Modern e-commerce search engines primarily rely on a passive "retrieval-and-ranking" paradigm. While effective for simple keyword matching, this approach fails in complex decision-making scenarios where users face:

Cognitive Friction: Users are overwhelmed by excessive information and comparison costs.
Passive Response: Systems act as "result displays" rather than "decision facilitators," unable to handle implicit, multi-attribute reasoning needs.
Data Silos: Reliance on internal platform data limits the integration of cross-platform knowledge (e.g., expert reviews, external guides).
Static Indexing: Traditional systems struggle to bridge the gap between fragmented product attributes and dynamic user intents during a dialogue.

The core challenge is transforming search from a static keyword match into an active, collaborative decision support system that mimics human cognitive workflows.

2. Methodology: The CogSearch Framework

CogSearch reimagines e-commerce search as a cognitive decision-making process inspired by Simon's three-stage decision model (Intelligence, Design, Choice). It employs a multi-agent framework consisting of four specialized agents and a central Memory System, orchestrated to guide the user from "need articulation" to "decision formation."

A. System Architecture

The framework operates through a collaborative loop of four agents:

Planner Agent (Intent Decomposition):
- Role: Acts as the cognitive orchestration engine. It bridges the semantic gap between ambiguous user queries and system execution.
- Mechanism: Takes the current query, search history, and click context ( $C_t$ ) and generates a Directed Acyclic Graph (DAG) of atomic tasks.
- Task Types:
  - ProductSearch: Retrieving candidates from internal inventory.
  - WebSearch: Acquiring open-world knowledge (reviews, guides).
  - ToolInvocation: Executing functional operations (e.g., price comparison, logistics) via Model Context Protocol (MCP).
Executor Agent (Multi-Source Acquisition):
- Role: Gathers heterogeneous information to feed the decision process.
- Components:
  - ProductSearch Agent: Uses a hybrid retrieval architecture (Keyword-Vector Fusion) and aggregates product attributes/reviews.
  - WebSearch Agent: Uses an "Expand-Retrieve-Filter" pipeline to fetch and score external data based on relevance, authority, and freshness.
  - Tool Integration Agent: Dynamically invokes external APIs (e.g., weather impact on logistics) based on context triggers.
Guider Agent (Proactive Guidance):
- Role: Solves the "paradox of choice" by providing dynamic, personalized guidance.
- Mechanisms:
  - Dynamic Cognitive Facet Generation: Uses a Clarify-Guide-Contextualize (CGC) mechanism to generate hierarchical filters based on the semantic variance between retrieved items and user state, reducing cognitive load.
  - Personalized Purchase Strategy: Synthesizes specs, web knowledge, and user history to create "Buying Guides" with trade-off analysis.
  - Query Suggestions: Proactively steers the search by suggesting Convergent Queries (zooming in on attributes) or Stimulative Queries (expanding based on latent needs).
Decider Agent (Integrative Reasoning):
- Role: The terminal cognitive nexus that outputs the final recommendation.
- Mechanism: Aggregates all upstream signals (products, web evidence, interaction history, user profile) into a unified context ( $S_{ctx}$ ).
- Decision Logic: Performs Multi-Criteria Decision Making (MCDM) using Large Language Models (LLMs). It evaluates candidates across four dimensions: Functional Performance, Economic Viability, Reliability, and Constraint Adherence.
- Output: A persuasive recommendation pair $(d^*, R)$ , where $d^*$ is the optimal item and $R$ is a natural language rationale explaining the decision.
Memory System:
- Acts as the central hub for storing contextual information, agent states, and the task graph to ensure persistent intent alignment across multi-turn interactions.

3. Key Contributions

Architectural Innovation: Proposes the first cognitive-aligned multi-agent framework for e-commerce search, shifting from "keyword matching" to "active decision assistance" via the Planner-Executor-Guider-Decider loop.
Cognitive Alignment: Explicitly models the user's shopping cognitive chain, using dynamic filtering and proactive guidance to reduce cognitive load and decision costs.
Cross-Source Fusion: Integrates internal platform data with external web knowledge and tool-based data to overcome information silos.
Industrial Validation: Successfully deployed on JD.com, one of China's largest e-commerce platforms, providing real-world validation of the framework's efficacy.

4. Experimental Results

Offline Evaluation (ECCD-Bench)

A new benchmark, ECCD-Bench, was created using 10k expert-annotated query-response pairs from JD.com.

Performance: CogSearch significantly outperformed the baseline (LLM-enhanced retrieval-ranking) in Complex Needs and Consultative Queries.
- Complex Needs: ACC@5 improved from 0.48 to 0.83; User Demand Satisfaction (UDS) improved from 1.92 to 2.79.
- Consultative Queries: ACC@5 improved from 0.60 to 0.65; UDS improved from 1.81 to 2.88.
Ablation Studies: Removing any single agent (especially the Planner or Guider) caused significant drops in performance, confirming the necessity of the full multi-agent collaboration.

Online A/B Testing (JD.com Deployment)

A two-week test with 10% of live traffic compared CogSearch against traditional search.

Decision Cost (DC): Reduced by 5%, indicating users reached decisions with fewer interactions (searches/clicks).
User Conversion Rate (UCVR): Increased by 0.41% overall.
Decision-Heavy Queries: For complex intents (e.g., gift selection, professional comparisons), UCVR surged by 30%.
User Behavior: The proportion of long queries (>10 words) increased by 25%, indicating a shift toward natural language expression and deeper engagement.
Latency: Despite an average end-to-end latency of 1.3s (vs. 800ms for traditional), streaming output maintained a responsive user experience.

5. Significance

CogSearch represents a fundamental paradigm shift in information retrieval:

From Passive to Proactive: It moves beyond simply returning relevant items to actively guiding users through complex decision trees.
From Relevance to Decision Intelligence: It prioritizes "decision quality" and "cognitive closure" over simple ranking scores.
Scalability: The successful deployment on a massive platform like JD.com proves that multi-agent LLM systems can be optimized for industrial-scale latency and reliability.
Future Direction: It lays the groundwork for "lifelong, multimodal shopping companions" that integrate sensory perception and long-term memory, evolving search into a holistic collaborative intelligence.