Product Search Semantic Ranking
Traditional product search mainly checks whether query terms appear in a title or attribute. That works for a precise brand or model number, but it struggles with conversational wording, synonyms, and broad needs. A query such as “a lightweight laptop for commuting” describes a situation and product qualities rather than an exact title.
This project replaced single-stage keyword matching with multi-stage ranking. It first found a broad set of plausible products, then progressively added semantic, behavioural, and business evidence. The aim was to improve relevance and conversion without sacrificing response time.
Role and key responsibilities
I led the ranking design and core implementation, with the wider search team delivering the production system. My responsibilities included:
- designing a retrieval, pre-ranking, and final-ranking pipeline so expensive processing was limited to a small candidate set;
- combining BM25 textual relevance with Embedding/BERT semantic representations;
- incorporating more than 20 signals, including clicks, sales, location, category preference, and product content;
- defining offline and online measures and building A/B testing, metric tracking, canary release, and rollback;
- controlling feature-computation, model, and network costs on the critical query path.
How the ranking pipeline worked
The first stage prioritised coverage. Keywords, category data, and semantic vectors produced a broad candidate set. The second stage applied moderately expensive signals to remove weak candidates. The final stage combined richer user and product evidence to determine the order shown to the customer.
This structure separated quality from cost. Expensive methods processed only a small number of candidates, while each stage exposed its own inputs, outputs, and latency. The team could distinguish poor recall from faulty features or ranking-policy regressions.
Controlled online validation
A few hand-picked queries cannot establish ranking quality. New strategies therefore received a small portion of live traffic first. We compared click and conversion behaviour, zero-result rate, latency, and errors with the existing strategy, then expanded traffic only when the benefit remained stable and important categories did not regress.
Experiment configuration, metric definitions, and rollback criteria were part of the delivery process. Algorithm changes became comparable and auditable engineering releases rather than opaque model replacements.
Outcomes
- The core path remained below 150 ms at P99 while evaluating more than 20 textual, semantic, behavioural, and business features.
- Search conversion increased by approximately 18% after release.
- Continuous A/B validation, canary rollout, and rapid rollback became standard capabilities for later ranking iterations.
The central achievement was placing stronger semantic understanding inside a production architecture that could preserve latency, prove business value, and control release risk.