Product Search RAG

A Streamlit app comparing keyword, semantic, and hybrid product search, with RAG answers from product reviews.

Live App View on GitHub Data on Hugging Face

The live app is on a free tier. If it has been idle, it can take 1 to 2 minutes to start.

The app may be asleep

This app runs on Streamlit Community Cloud's free tier, which puts it to sleep after a period of no use.

If you see a page saying the app has gone to sleep, click "Yes, get this app back up!" and give it 1 to 2 minutes to start.

Open the app

This project is a product search assistant built on the Appliances category of the Amazon Reviews 2023 dataset from McAuley Lab. It compares keyword, semantic, and hybrid retrieval side by side, and adds two RAG modes that answer natural language questions using retrieved product details and reviews.

It started as a team project by Christine Chow and Ian Gault for DSCI 575 (Advanced Machine Learning) in the UBC Master of Data Science program. The public repository is my continuation of that work.

What it does

• BM25 search - keyword ranking with rank_bm25 over product titles, features, descriptions, categories, details, and review text
• Semantic search - all-MiniLM-L6-v2 sentence embeddings indexed with FAISS, matching queries by meaning rather than exact words
• Hybrid search - combines BM25 and semantic results with Reciprocal Rank Fusion
• RAG and Hybrid RAG - retrieved products and review text are passed as context to a Qwen model via the Groq API, which writes a grounded answer shown alongside the supporting products
• Query guardrail - an LLM-based check rejects questions unrelated to appliances or product shopping

Retrieval evaluation

Each retriever was scored on 10 test queries against human relevance labels (k = 5):

Retriever	Mean Precision@5	Mean Recall@5	MRR
BM25	0.54	0.33	0.73
Semantic	0.68	0.39	0.85
Hybrid	0.58	0.36	0.85

Semantic search scored highest on this query set and tied with hybrid search on MRR. Some queries were written to favour semantic interpretation (for example, misspellings), which may explain part of the gap.

Technical stack

Area	Tools
Data processing	DuckDB, pandas, NLTK
Keyword retrieval	rank_bm25
Semantic retrieval	sentence-transformers, FAISS
RAG	LangChain, Groq API (Qwen)
App framework	Streamlit
Data hosting	Hugging Face Hub
Deployment	Streamlit Community Cloud
Reproducibility	conda, GNU Make

What I changed for the public version

• Reworked the project into a standalone repository with a new README
• Switched the LLM to qwen/qwen3.8-27b after Groq retired the original qwen/qwen3-32b
• Moved the processed data and search indexes (about 207 MB) to a public Hugging Face dataset, so the git repository stays small and the app downloads them at startup
• Deployed the app to Streamlit Community Cloud