Neural E-Commerce Search

Neural E-Commerce Search is an alpha research implementation of a two-stage product-search pipeline on the Amazon ESCI benchmark.

query -> bi-encoder -> candidate set -> DeBERTa cross-encoder -> ranked ESCI labels

Try neural retrieval in the browser

The NECS Browser Lab runs a public MiniLM encoder through Transformers.js and ranks a small synthetic product catalogue entirely on the visitor’s device. It needs no account or local environment and makes its model, scores, and evidence boundary visible.

The lab is a real dense-retrieval demonstration, but it is not the repository’s unreleased ESCI-trained bi-encoder/cross-encoder pipeline and does not reproduce historical benchmark figures.

Start without data or models

From a source checkout, run the deterministic synthetic demo:

python scripts/offline_demo.py --query "wireless gaming mouse"

It demonstrates the output contract with transparent heuristics. It does not run the neural models and is not benchmark evidence.

Evidence status

The code, configs, lightweight tests, browser retriever, and offline demo are public. ESCI-trained project weights, full raw benchmark artifacts, multi-seed reruns, and a hosted end-to-end reranker are not yet published. Historical metrics are withdrawn pending a corrected rerun. Read Experiments and Reproducibility before citing them.

Documentation

• Architecture: two-stage design and module map
• Data: ESCI taxonomy, provenance, and preprocessing
• Training: training workflow and configuration
• Experiments: evidence status and rerun plan
• Reproducibility: publication checklist
• Deployment: serving and operational caveats
• Browser lab: client-side MiniLM retrieval demo
• Releasing: package and release safety checks

Links

• Source code
• Roadmap
• Contributing

MIT licensed. Built on the Amazon ESCI benchmark (Reddy et al., 2022).