Search Functionality for a Ecommerce site.

• Try it out here: https://ecommerce-retrieval-search.streamlit.app/
• View the doc and approach for this project here: https://1drv.ms/w/s!AnO5FdGErMSuh7VXK1wgpzwajPXOGw?e=fAh5Qh

Overview

This project aims to enhance the search experience for an e-commerce platform, by implementing an efficient and accurate product retrieval system. The system uses natural language processing techniques and vector similarity search to match user queries with relevant products.

Features

• Text-based product search
• Image retrieval based on text queries
• Utilizes product metadata for improved search accuracy
• Fast similarity search using FAISS (Facebook AI Similarity Search)
• Simple and intuitive web interface

How it works

1. Data Preprocessing:

   • Cleans and combines relevant product information (name, description, category, specifications, brand)
   • Performs text preprocessing (removing HTML tags, lowercasing, removing punctuation and stopwords)

2. Embedding Generation:

   • Uses the 'gte-base-1.5' embedding model (768 dimensions)
   • Converts preprocessed text into dense vector representations

3. Similarity Search:

   • Utilizes FAISS for efficient storage and querying of embeddings
   • Implements L2 distance (Euclidean) for similarity measurement

4. Query Processing:

   • Applies the same preprocessing to user queries
   • Generates embeddings for the query
   • Retrieves top K most similar products using FAISS

5. Result Display:

   • Fetches and displays product images based on the retrieved results

Tech Stack

• python: Primary programming language
• pandas: Data manipulation and CSV handling
• matplotlib: Data visualization
• transformers: Embedding model
• pyTorch: Tensor operations and GPU support
• nltk: Text preprocessing (stopwords, stemming)
• faiss-cpu: Vector store for similarity search
• streamlit: GUI and deployment

Installation and Usage

• Clone the repository.
• Install all the dependencies from requirements.txt file. Run !pip install -r requirements.txt in the terminal.
• Run streamlit run app.py and the app will run on localhost.

Future Improvements

• Implement multimodal search using vision-language models like CLIP
• Enable image-to-image and image-to-text queries
• Fine-tune image models (e.g., ResNet, ViT) on the product dataset
• Implement re-ranking for multimodal queries
• Create a manually curated test set for evaluation (using metrics like Recall@K)
• Integrate an LLM for handling malformed and multilingual queries

Results

• Query: A Red skirt
• Query: Football shoes
• Query: Running shoes
• Query: Superhero t-shirt