Vektor Commerce
Intelligent Search
A hybrid vector search engine for a 500k-SKU catalogue — replacing keyword search with semantic understanding that surfaces the right product even when customers don't know the exact name.
The problem: search that couldn't understand customers
Vektor ran a B2B industrial supply catalogue with 500,000+ SKUs. Their keyword search worked when buyers knew the exact part number. When they didn't — when they searched "something to seal pipe threads in cold environments" — it returned nothing, or worse, unrelated results.
Internal data showed 38% of searches ended with no click. Their sales team was fielding calls from buyers who couldn't find what they needed online. The catalogue existed, but was effectively invisible.
"We had the inventory. The problem was customers couldn't find it. Half our support tickets were literally just 'help me find a product'."
— VP of Digital, Vektor CommerceWhat we built
We replaced the keyword engine with a hybrid search architecture: BM25 for exact-match retrieval (part numbers, brand names) layered with dense vector search for semantic queries. Every product was embedded using OpenAI's text-embedding-3-large model and indexed in Pinecone.
We also built an AI catalogue tagger — a background pipeline that ran against the entire SKU database, enriching sparse product descriptions with standardised attributes: material, temperature range, industry application, compatible fittings. This dramatically improved both search relevance and filter accuracy.
Catalogue audit & embedding pipeline
Assessed data quality across 500k SKUs. Built an async embedding pipeline — chunking, cleaning, and indexing into Pinecone in 72 hours. Identified 180k products with sparse descriptions as priority for AI enrichment.
Hybrid retrieval engine
Built the BM25 + vector fusion layer. Tuned RRF (Reciprocal Rank Fusion) weighting based on a test set of 2,000 representative customer queries, evaluated by the Vektor sales team.
AI catalogue tagger
Designed a LangChain enrichment pipeline. Ran overnight on the 180k sparse SKUs. Generated structured attribute tags: environment rating, material family, application category. Cut manual tagging backlog from 18 months to zero.
Next.js frontend & A/B rollout
New search UI with faceted filtering on AI-generated attributes. A/B tested against old search over 3 weeks. New system showed +34% conversion and −61% in zero-result searches.
Impact after 3-week A/B test
The A/B test ran across 120k unique sessions. The new search system outperformed keyword search on every metric — most significantly, zero-result searches dropped from 38% to under 15%.