README.md

metadata

title: ShopRLVE Cart Environment
emoji: 🛒
colorFrom: green
colorTo: blue
sdk: gradio
sdk_version: 6.0.0
app_file: app.py
python_version: '3.12'
pinned: false
license: apache-2.0

ShopRLVE-GYM — Cart Environment Demo

Interactive demo for the Cart Building environment from ShopRLVE-GYM.

You play the AI agent. A persona-driven simulated customer asks you to find and cart specific products. Use catalog search, cart tools, and chat to fulfill the request, then submit your answer to see the verifiable reward breakdown.

How it works

1. Reset an episode at your chosen difficulty (0-10)
2. Read the customer's request in the chat
3. Use tools (search, get_product, get_variants, cart.add, etc.)
4. Chat with the customer for clarification
5. Submit your answer with the product IDs in the cart
6. See the reward decomposition: r_task (75%), r_eff (15%), r_hall (10%)

Data requirements

Primary: Real Catalog (recommended)

The app prefers a 29.6K stratified real product catalog with a pre-built FAISS index using Alibaba-NLP/gte-modernbert-base (768-dim) embeddings.

Required files in data/real_catalog_index/:

File	Size	Description
catalog.jsonl	7.9 MB	29,601 products (JSONL: id, title, category, brand, price, rating)
index.faiss	87 MB	FAISS Flat index (768-dim, inner product)
ids.txt	318 KB	Product ID list matching index order
meta.json	138 B	Index metadata

Build with: python scripts/build_real_catalog_index.py

Source: data/real_product_catalog_stratified.jsonl (29.6K products stratified from 231K Amazebay catalog — 5 products per product type across 6000 categories).

Fallback: HuggingFace Dataset

If the real catalog is not found, loads from thebajajra/Amazebay-catalog on HuggingFace Hub (or a local Arrow dataset), limited to 5000 items. Builds a FAISS Flat index at startup (~30s on CPU).

Environment variables

Variable	Default	Description
REAL_CATALOG_PATH	data/real_catalog_index/catalog.jsonl	Real catalog JSONL
REAL_INDEX_DIR	data/real_catalog_index	Directory with index.faiss + ids.txt
CATALOG_PATH	data/amazebay-2M	Fallback: HF dataset path
CATALOG_MAX_ITEMS	5000	Fallback: products to load
EMBEDDING_MODEL	Alibaba-NLP/gte-modernbert-base	Must match index embeddings
EMBEDDING_DEVICE	cpu	Device for query encoding