How a DAG‑Based Word Distribution Platform Boosts E‑Commerce Search Efficiency

Background

Word distribution (导购) guides users during the shopping process by recommending hot items, promotions, or personalized content, thereby improving the shopping experience and driving sales.

Key Component: Word Distribution

Word distribution focuses on keyword‑related functions such as dropdown suggestions, background words, hot‑search lists, tip words, and trend indicators, helping users quickly locate products and reducing search friction.

Word Distribution Platform

The platform creates a unified, efficient ecosystem by integrating multiple algorithms, tools, and generic service interfaces, offering several advantages:

Unified development, reduced cost : Standardized services and interfaces avoid duplicate development across business domains.

High flexibility, multi‑scenario support : Modular design adapts quickly to promotions, holidays, or category‑specific recommendations.

Supports business expansion : Teams can focus on core logic while the platform handles recommendation infrastructure.

Optimized user experience : Advanced algorithms and data analysis deliver precise, personalized suggestions.

Supported Scenarios

More than 30 community, transaction, and marketing scenarios are already supported.

Overall Architecture

Business Architecture

The business layer defines the end‑to‑end flow (pre‑processing → recall → fusion → coarse ranking → fine ranking → result packaging) and abstracts each stage as a configurable component managed by a central orchestration engine.

Platform Architecture

The platform layer provides a Spring‑based container, a core word‑distribution engine, dependency injection, and logging/debug capabilities, enabling scalable and maintainable services.

Architecture Evolution 3.0 – Graphization

Serial Architecture

Initially, the system followed a linear processing pipeline.

Graph Engine Evolution

To address flexibility and performance challenges, the architecture migrated to a directed‑acyclic‑graph (DAG) model where nodes represent operators and edges represent data flow. This enables unified operator interfaces, reusable components, parallel and asynchronous execution, and easier experimentation.

Core Design and Protocol

Operators implement the IDagTaskNodeExecutor<T> interface and are annotated with @DagNodeMetaProcessor to declare name and description. The DAG engine parses the logical graph, schedules operators, and ensures stability and performance.

@Inherited
@Retention(RetentionPolicy.RUNTIME)
@Target(ElementType.TYPE)
public @interface DagNodeMetaProcessor {
    /** Operator name */
    String name();
    /** Operator description */
    String desc() default "";
}

public interface IDagTaskNodeExecutor<T> {
    T execute(DagStrategyContext dagStrategyContext);
}

Script Hot‑Deployment Architecture

The hot‑deployment mechanism allows dynamic loading of strategy scripts without restarting services. Developers define a unified abstract method, implement the logic in an SDK, and the backend packages, configures, and pushes the code. At runtime the service reflects the new class and executes it based on AB testing configuration, dramatically reducing iteration time and maintenance cost.

Main vs. Script Projects

Main project base : Provides abstract and generic layers, ensuring a stable framework and reusable functionality.

Script project : Focuses on business‑specific logic, allowing rapid adaptation to diverse scenarios while leveraging the base framework.

Supporting Tools

Scripted Development & CI/CD : Automated build, gray‑release, and full‑release pipelines with version synchronization.

DIFF Evaluation Platform : Visual tool for assessing algorithmic changes.

Community Search Evaluation Platform : System for extracting query data, snapshots, exporting evaluation data, and analyzing results.

Intervention Platform : Enables real‑time control over search terms (e.g., hot‑event insertion, term deactivation).

Recall Configuration Platform : Visual management of multi‑recall configurations, replacing manual JSON edits.

Future Planning

Enhance platform monitoring, expand recall configuration and script CI/CD capabilities.

Upgrade the base framework to support Spring integration and improve maintainability.

Accelerate onboarding of new scenarios such as detailed‑page triggers and “small‑blue” terms.

Break data silos between product‑centric and term‑centric pipelines to boost intelligence.