Search & retrieval
- Input → outputTexttoRanked results
- ModelLiquid 350M + ColBERT
- Runs onCloud GPU
- What it doesIntent · expand · rescue · re-rank
Recoverreal intent.
Search intelligence that connects shopper intent to your catalog.
Searching Cedar Supply Co., a demo store with 100 products in 16 categories.
Classify the query before it reaches the index.
Problem
Shoppers describe what they need, not how your catalog names it. A query like ‘cozy blanket’ can miss a Fleece Sherpa Blanket when the words don’t match. Ambiguous intent, empty results, and poorly ordered matches interrupt the journey from search to purchase. An empty page might mean the query needs help, or that the product simply isn’t stocked.
Solution
A fine-tuned 350M Liquid model classifies intent, expands queries, rescues phrasing mismatches, and re-ranks results. A ColBERT retriever matches query and product tokens. Products stay in the index, separate from the models, so catalog updates don’t require retraining. The demo lets you inspect each stage against the same catalog, including searches it cannot recover.
Match what shoppers mean.
Bridge everyday descriptions and catalog terminology before a query reaches the index.
Recover the right searches.
Separate a phrasing mismatch from a product you don’t stock. Rewrite when it can help.
Keep your catalog independent.
Update the product index as inventory changes. The query specialist stays the same.