E-commerce2023Confidential client

AI product search and recommendations

Search that understands what shoppers mean, not just the words they type, plus recommendations built from the catalogue and real browsing behaviour.

LLM Solutions
Software Development

Start a similar project All work

Client: Confidential client
Industry: E-commerce
Year: 2023
Services: LLM Solutions · Software Development

The challenge

Keyword search missed what shoppers meant, so good products went unfound and searches ended with no result.

The store's keyword search failed on the queries shoppers actually type: "something warm for a winter wedding", misspellings, Roman Urdu, and product names that differ from how customers describe them. Too many searches returned nothing, and the merchandising team was hand-curating "you may also like" rows.

What we built

Semantic search over the catalogue with typo and intent handling, plus recommendations from browsing behaviour.

Hybrid search
Keyword matching (still the best for exact SKUs and brands) combined with semantic vector search over product titles, descriptions and attributes, then re-ranked for stock and margin rules the business controls.
Query understanding
An LLM step turns a long natural-language query into filters such as category, size, colour and price range, which shoppers can see and remove. Nothing is hidden behind a black box.
Recommendations
Built from co-viewing and co-purchase patterns, with sensible fallbacks for new products that have no history yet.
A merchandiser console
Merchandisers pin, boost and block results per query, with a preview before changes go live.

How we measured it

Before switching traffic, we replayed real anonymised search logs against both engines and had the merchandising team judge the top results side by side. The new search went live behind a feature flag for a share of visitors first.

Outcome

Shoppers get relevant results for descriptive and misspelt queries that used to return an empty page, and the merchandising team tunes search themselves instead of filing tickets.