All case studies / Case study 06 / 12
Kaldo Sport
Conversational shopping assistant
A sales adviser that understands what customers really want
In short
Customers describe their need as they would to a shop assistant. The assistant turns it into precise criteria, finds the products that truly match (specs, descriptions, reviews) and explains its choice.
- Client
- Kaldo Sport
- Sector
- Outdoor e-commerce
- Size
- 8,000 products, 40,000 visits a month
- +24%conversion on assisted sessions
- −35%returns for “not suitable”
- 24/7in French, Arabic and English
The context
Kaldo Sport sells hiking and camping gear online. Its customers know what they want to do, “three days in the mountains in February”, but not which product to pick among hundreds of items with complex specs.
The challenge
- 01
Classic filters (price, brand) don’t understand “waterproof for three days in the cold”.
- 02
Keyword search misses products described differently.
- 03
Recommendations must respect real stock, prices and specs: nothing made up.
- 04
The assistant must plug into the existing shop, not replace it.
Our approach
Every product page, its specs and its reviews are indexed in a vector database with their metadata: price, stock, season, waterproofing, weight. The customer’s request is split into structured filters and intent, the search combines both to find the best products, and the model writes a recommendation based only on those pages.
The solution, step by step
- 01
The need, in plain words
“A waterproof tent for 3 days in the mountains in February, under TND 600.”
- 02
Turned into criteria
Waterproof, cold season, 2 people, light, max price: extracted automatically.
- 03
Filtered, then ranked
The catalogue is filtered on those criteria, then ranked using descriptions and reviews.
- 04
The best, in stock
Only available, matching products remain, from best to least suited.
- 05
An explained recommendation
The assistant explains its pick, flags trade-offs and suggests matching gear.
Results
- +24%conversion on assisted sessions
- +18%average basket thanks to related products
- −35%returns for “not suitable”
- < 2 sper recommendation
What we deliver
- Indexing of catalogue, product pages and reviews
- Criteria extraction and hybrid search with metadata
- Chat widget built into the shop
- Stock and price sync
- Conversion and frequent-question tracking
Technologies
- Python
- LlamaIndex
- Pinecone
- React
- Shop API
- Docker
Timeline
6 to 8 weeks
Team
1 AI engineer, 1 front-end developer
Illustrative cases: clients and figures are fictional, representative of the solutions we deliver.