Software · Data and AI
Search and recommendations
Search that finds the right item from vague queries, and recommendations based on what people really browse and buy.
When search returns nothing for a misspelt product, the customer leaves. We build search that copes with typos, synonyms, units and mixed scripts, combining keyword and semantic matching, then ranks results using your business rules and real behaviour. Recommendations are built the same way, and every ranking change is A/B tested before it reaches everyone.
What we build
- Product and content search with typo tolerance, synonyms, filters and facets
- Hybrid keyword and vector search for vague or descriptive queries
- Recommendation widgets for similar items, bought together and recently viewed
- Merchandising controls to pin, boost or hide results for specific queries
- Search analytics showing zero-result queries, clicks and refinements
- A/B testing of ranking changes on a share of traffic before full rollout
DemoInteractive · sample data
Try it live.
A working sample with made-up data. It runs in your browser; nothing is sent anywhere.
{
"index": "products",
"query": "warm jaket for trekking",
"mode": "hybrid",
"weights": { "keyword": 0.4, "vector": 0.6 },
"typo_tolerance": true,
"filters": { "in_stock": true, "size": "M" },
"boost": { "rating": 1.2 },
"limit": 5What it does well
Typo and synonym aware
"Iphon caes" finds iPhone cases, and a search for sofa also matches couch and settee.
Meaning-based matching
Vector search finds a "warm jacket for trekking" even when no product uses those words.
Your ranking rules
Stock, margin, freshness and labelled sponsored placements are weighed with relevance.
Zero-result fixes
A report of failed searches shows which synonyms or products to add next.
Tested changes
Ranking changes run as A/B tests and are kept only if clicks and orders improve.
Live index
Indexes update as stock and prices change, so sold-out items drop out of results quickly.
Questions
Should we use a hosted search service or build our own?
Hosted search services are quick to launch and fine for many catalogues. A custom setup makes sense when you need control over ranking, have unusual data or languages, or when per-search pricing grows too costly. We compare both for your traffic and catalogue.
How do recommendations start with no history?
New products and new visitors have no behaviour to learn from, so we begin with content similarity: category, attributes, descriptions and images. As views and purchases build up, behaviour-based signals gradually take over.
How do we know search got better?
We track zero-result rate, clicks on the top results, search-to-order conversion and refinements, before and after each change. Ranking changes are A/B tested on part of the traffic, so you decide on measured results rather than opinion.