README.md

---
license: odc-by
language:
  - en
size_categories:
  - 1M<n<10M
dataset_info:
  features:
    - name: uid
      dtype: string
    - name: vendor
      dtype: string
    - name: title
      dtype: string
    - name: paragraph
      dtype: string
    - name: embedding
      dtype:
        sequence: float32
task_categories:
  - text-retrieval
  - sentence-similarity
task_ids:
  - document-retrieval    
  -  semantic-similarity-classification
tags:
  - ecommerce
  - small-business
  - rag
  - grounding
  - vector-search
  - open-data
  - embedding
  - tokuhn
  - shopify
  - real-world-data
  - sbert
  - huggingface-datasets
---
# [Updated with SBERT Embeddings + Search Notebook] 
## TSMPD‑US: U.S. Small Merchant Product Dataset + SBERT Embeddings + Search Notebook

⚡ New in this release (April 2025):

SBERT vector embeddings for all products (MiniLM‑L6)

Chunked Parquet format for scalable vector search

Jupyter notebook demo for live semantic queries

These additions make it easier to integrate small merchant data into RAG pipelines, grounding tasks, and real-time AI agents.

## An open-source initiative to keep small merchants visible in LLMs, RAG systems, and AI-powered commerce workflows.**

This repository contains multiple assets for the TSMPDUS dataset  a structured, U.S.-only dataset of verified small business product listings, curated from over **355,000 independent stores**. It is designed for:

- Semantic product search
- LLM grounding and fine-tuning
- Retrieval-Augmented Generation (RAG)
- Metadata classification
- Commerce-aware agent design

---

## Directory Overview

### `public-products/`

A lightweight, text-only snapshot of the dataset.
- **~3.2M products** from 355,000+ verified U.S. merchants
- ~10 products per merchant, no images or variant details
- Suitable for general research, classification, and basic grounding tasks

**Includes:**
- `tsmpd_public_v1.0.json` or `.parquet`  core dataset
- `LICENSE.txt`  ODC-By license
- `README.md`  Format & schema details

---

### `parquet-embeddings/`

Semantic searchready version of the dataset with **SBERT embeddings** (MiniLML6).
- Split into Parquet chunks for scalability
- Embeddings aligned with Hugging Face `sentence-transformers/all-MiniLM-L6-v2`

**Use cases:**
- Vector search & similarity pipelines
- Retrieval-Augmented Generation (RAG)
- AI agent product reasoning

**Includes:**
- `tsmpd_public_000.parquet`, `...001.parquet`, etc.
- `README.md`  Usage notes + embedding shape
- `LICENSE.txt`  Same ODC-By license unless extended

---

### `notebook-demo/`

A minimal working demo for semantic product search over the embedded dataset.
- Loads Parquet embeddings
- Performs cosine similarity on live queries
- Displays top product hits from the network

**Includes:**
- `tsmpd_search_demo.ipynb`  Search notebook
- `README.md`  Instructions & dependencies

---

## Why This Matters

Large models like ChatGPT and Claude do not crawl small stores the way Google does. Without structured visibility, the **long tail of small commerce risks becoming invisible** in AI-powered discovery systems.

**TSMPD-US** is designed to prevent that  by making small merchant data accessible, embeddable, and usable in todays LLM workflows.

---

## Licensing

All public assets are distributed under the [Open Data Commons Attribution License (ODCBy)](https://opendatacommons.org/licenses/by/1-0/).

For full product variants, image URLs, merchant domains, and source tracking, request access to the **Partner dataset** by emailing `jim@tokuhn.com`.

---

## How to Use This Repository

- Load the text-only dataset via Hugging Face Datasets or `polars`
- Run vector search with the SBERT Parquet chunks
- Adapt the notebook demo for your own semantic or retrieval tasks
- Fine-tune or evaluate grounding quality with real-world small merchant data

Lets make sure AI doesnt erase the 99%.

---