E‑commerce Search Engine Recall and Vector Retrieval Techniques

Search engines are widely used in e‑commerce; to retrieve items efficiently among billions of products, an inverted index is built so that each searchable token maps to the products containing it.

The first stage, called recall, quickly narrows the candidate set using simple matching before more expensive ranking stages.

Word‑based recall relies on proper tokenization and query rewriting (e.g., “圆领” AND “T恤”) to intersect posting lists and obtain relevant items, while a recall score limits the number of candidates passed to later stages.

To address category bias and improve relevance, recall scores consider product quality, release time, and can be adjusted per‑category using predictions.

Vector‑based recall complements word recall by capturing semantic similarity; the problem is formulated as a classification task with cross‑entropy loss, and pairwise or triplet losses are also described.

Deep learning models such as Microsoft’s DSSM (Deep Structured Semantic Model) and Google’s DeepMatch embed queries and items into vectors and compute cosine similarity for semantic matching.

A personalized real‑time recall model combines offline‑precomputed user ID features with an online attention network that processes the user's recent behavior sequence, producing a real‑time user vector for fast Top‑N retrieval.

Additional recall paths include I2I expansion from items the user has interacted with, and text‑based recall using DSSM on product titles and copy, as well as multi‑path fusion to enrich candidate sets.

The system is deployed by splitting the model into offline and online components to accelerate inference, and the overall loss remains cross‑entropy with importance sampling of positive and negative samples.