Embed
MiniLM converts query and product text into normalized vectors.
Runs locally in your browser
Search products with real neural embeddings.
Type a shopping query. A compact MiniLM model embeds the query and catalogue on your device, then ranks products by semantic similarity. There is no project backend, account, or uploaded query; model and library assets are fetched from Hugging Face and jsDelivr.
Client-side inference
No signup
Open source
20-product sample
Try the retriever
What are you looking for?
Model loads on first search
Try:
Loading model assets
0%
The first query downloads roughly 25 MB of quantized model assets, plus the Transformers.js library, then caches them. Later searches are fast.
Retrieve
Cosine similarity selects the closest products in embedding space.
Inspect
Scores and product metadata stay visible instead of hiding the result.
Ranked catalogue
Results
Run a search to populate the ranking.
The model is ready when you are.
Your query is processed on this device.
Evidence boundary
A real neural demo, with a precise claim.
This lab uses the public
all-MiniLM-L6-v2
encoder at revision 751bff3 through Transformers.js 3.8.1
for dense retrieval over a synthetic catalogue. It does not run this
repository's unreleased ESCI-trained bi-encoder or DeBERTa reranker,
and it is not benchmark evidence. The distinction is intentional.
Build or break it
Useful for your search work?
Replace one JSON file to search your own catalogue, reproduce a result, or report the first confusing edge case.