This is a simple Node.js application that classify e-commerce products into the categories of well-known e-commerce websites like Amazon, Shopify and Google Shopping. It supports multiple languages.

It depends on OpenAI Embeddings API to get the embeddings of categories and products. And it uses Langchain/HNSWLib to build a fast and scalable search index.

Endpoints

There are two endpoints mainly. One is to create vectors and the other is for classifying a product into a category. One is based on Pinecone to keep the vectors and the other is based on HNSWLib to keep the vectors locally. Both of them use OpenAI Embeddings API.

HNSW-based (recommended, faster)

• POST /v2/product/predict/:platform - creates vectors of categories for the given input file
• POST /v2/product/:platform - returns most-related category for the given product name

Pinecone-based (legacy, slower)

• POST /category/predict/:platform - creates vectors of categories for the given input file
• POST /category/:platform - returns most-related category for the given product name

How to classify new platform categories

First, you need to prepare a list of categories of the platform that you wanted to classify. For instance, this file is the product taxonomy file of Shopify. You can use it to create a list of categories. The input file should be TSV format, and the first column should be the ID and the second column should be the category name. You can find a sample input file under sample folder.

You can also find Postman collection in postman.json file.

Getting Started

Installation

npm install

Development

To run the application in development mode:

npm start

Production

To run the application in production mode:

npm run build
npm run start:live

Bonus: OpenAI ChatGPT Plugin

This repository also includes a plugin for OpenAI ChatGPT. You can use it to generate a response for a given message. You can find the plugin under static folder. If you have access to ChatGPT Plugins, you can load it from localhost:6006 and use it by asking a question like What is the best category for a product with name "iPhone 12 Pro Max"?. Do not forget to run the application. The plugin version uses the HNSW-based endpoint (/v2).

Todo

• Add different human-languages support at platform-level
• Remove pinecone-based endpoints once the HNSW-based endpoints are stable
• Prepare ready-to-use vectors for well-known platforms