Smarter search
using actual AI.
Semantic search powered by real AI embeddings.
Not keyword matching wearing a fake mustache.
The Problem
Your customers search "comfy blue sweater" and get zero results because no product has those exact words. They search "jean" and miss "jeans" and "denim."
Bad search = lost sales. Simple as that.
Key Features
AI Semantic Understanding
We use AI embeddings (384 dimensions) to understand what customers MEAN, not just what they TYPE. "Cozy throw blanket" finds your "soft knit blankets" even without keyword overlap.
Hybrid Search
We combine semantic AI with traditional keyword search using Reciprocal Rank Fusion. You get intent understanding AND exact matches. Best of both worlds.
Ecom Taxonomy Boost
Search "tops" and we expand to jackets, shirts, blouses, sweaters, hoodies. Automatic singular/plural handling. Category-aware ranking.
Quality Filtering
We filter out low-quality semantic matches when strong keyword matches exist.
Typo Tolerance
"Swaetr" → Sweater. "Runnig shoes" → Running shoes. Humans make typos. Your search shouldn't punish them.
Instant Results
Results appear as they type. No submit button. No page reload. Just fast.
Smart Synonyms
You define them or we learn them. Sofa = Couch. Sneakers = Tennis Shoes.
Search Analytics
See what people search for. See what gets zero results. See what converts. Fix the gaps.
Merchandising Controls
Boost products, bury products, add redirects and pages.
Technical Callouts
Your Data Stays Yours
Embeddings are generated and stored per-shop. Your product data doesn't train a shared model.
Voice Search
Use browser native voice transcription for quick and easy voice search.
Blazing Fast
Compression shrinks responses by 70%. Caching means repeat searches return instantly. Your customers get results in milliseconds.
How semantic search actually works
Products become vectors
Every product is converted into an embedding — a numeric representation of meaning — using an all-MiniLM-L6-v2 model. Two products end up close together when they mean similar things, not when they share a word.
Hybrid, not purely semantic
Pure semantic search is bad at exact matches like a SKU or a brand name. Queries run against both the embeddings and a keyword index, so "AirFlow 3000" still finds the AirFlow 3000.
You keep merchandising control
Relevance is a starting point, not a verdict. Specific queries can boost, pin or redirect results, so a search for a term you're running a campaign on lands where you want it.
AI search questions
Does this replace my theme's search?
It can take over the existing search field rather than adding a second one, so shoppers use the same box they always did and the theme's design is preserved.
What happens when someone misspells a product name?
Typo tolerance is configurable fuzzy matching rather than exact-string matching, and because the query is also interpreted semantically, a misspelling that still reads as the right concept returns the right products.
What about searches that return nothing?
Zero-result queries are recorded in search analytics. That list is the single most useful merchandising report you can read — it's your customers telling you, in their own words, what they expected you to sell.
Where do the embeddings run?
Product embeddings are generated for your catalog and used to serve your storefront's searches. Your catalog is not pooled with other merchants' to train a shared model.
