Design and Implementation of an E‑commerce Search Engine Using Go and Elasticsearch

The article begins by outlining the background of modern search technologies, noting that while Elasticsearch solves low‑level indexing, a production‑grade e‑commerce search engine still requires extensive engineering work.

It then describes the overall search architecture, including the core business flow (search entry → trigger → input → click → result) and presents diagrams of the technical stack: Go with the Kratos framework for backend services, Elasticsearch/OpenSearch for index storage, Redis for hot data caching, MySQL for persistence, and Kafka for message exchange.

Next, the article details the e‑commerce data model and provides a complete Elasticsearch mapping for product documents, showing fields such as id , spu_name , category_name , detail , price , and various status flags.

It discusses handling of bulk and incremental data ingestion, illustrating pipelines for historical data, incremental updates, and specialized indexing of product tags and promotional information with accompanying flowcharts.

The piece then examines Chinese tokenizers, comparing fuzzy, custom, pinyin, keyword, and domain‑specific analyzers, and recommends IK or Jieba for e‑commerce scenarios.

Following this, the design of pre‑search user‑guidance components—search watermark, history, and discovery—is presented, each with its own workflow diagram.

The core search service architecture is introduced as a three‑layer system: an Application Service (AS) layer written in Go, Elasticsearch for indexing, and a database layer. The AS layer is plugin‑based, allowing custom query rewriters, rerankers, and other algorithms to be added without changing the core.

Query processing (QP) is broken down into normalization, spell correction, tokenization, term weighting, stop‑word removal, and query rewriting. Sample query transformations are shown in code blocks, illustrating how stop‑words, spelling fixes, term weights, and synonyms affect the final query sent to Elasticsearch.

Ranking is described as a two‑stage funnel: a coarse ranking that filters large candidate sets using BM25‑like scores, followed by a fine‑grained ranking that incorporates user behavior features.

Finally, the article covers result presentation, including aggregation, filtering, and category faceting, and concludes with a discussion of future extensions such as image‑based search and recommendation systems that would integrate machine‑learning components.