daniloxvalle/esci-enriched-queries

ESCI Enriched Queries

Conversational rewrites of 1,061 shopping queries from the Amazon Shopping Queries Dataset (ESCI), with the relevance labels and the 3,121-product catalog they were evaluated on. Together, the three configs form a ready-to-use product retrieval benchmark with natural-language queries.

Why

ESCI queries are short keywords, often with typos ("johnnys chicken and pork seasoning"). Conversational recommender systems receive natural-language questions instead. Each query was rewritten as a short question a customer would ask a shop assistant ("Do you have Johnny's Chicken & Pork Seasoning?"), with the same intent and product requirements.

Configs

Config	Split	Rows	Content
queries (default)	test	1,061	Original and enriched queries
qrels	test	3,527	Relevance labels (query, product)
corpus	corpus	3,121	Product catalog searched by the retriever

Each config is in two files with the same rows: a Parquet file (loaded by the dataset viewer and load_dataset) and a JSON Lines file (one JSON object per line, for use without pandas).

queries

Column	Type	Description
query_id	int64	Query id in ESCI (shopping_queries_dataset_examples)
query	string	Original ESCI query, verbatim (typos included)
enriched_query	string	Conversational rewrite generated by OpenAI o3

query_id	query	enriched_query
396	0 cal water inhancer without sugar	Do you have sugar‑free, zero‑calorie water enhancers?
756	1 huba buba origanal not exspencive	Show me an inexpensive pack of original Hubba Bubba gum.
56184	johnnys chicken and pork seasoning	Do you have Johnny’s Chicken & Pork Seasoning?

qrels

Column	Type	Description
query_id	int64	Query id, as in queries
product_id	string	Product ASIN, as in corpus
esci_label	string	ESCI label: E (Exact), S (Substitute) or C (Complement)
gain	float64	Graded relevance used for nDCG: E = 1.0, S = 0.1, C = 0.01

The labels are the original ESCI annotations: 2,489 Exact, 1,010 Substitute and 28 Complement. Irrelevant (I) pairs are not included; a product with no row for a query has gain 0. Every query has 2 to 19 labeled products.

corpus

Column	Type	Description
product_id	string	Product ASIN (ESCI product_id, ESCI-S asin)
title	string	Product title from ESCI-S
description	string	Product description from ESCI-S, 1,500 to 2,000 characters
bullets	list of string	Product bullet points from ESCI-S

643 of the 3,121 products have no label for any query. They are part of the catalog the retriever searches, so the corpus cannot be rebuilt from qrels alone.

The text fields are copied verbatim from ESCI-S: 2,976 descriptions begin with the literal header "Product Description", and 4 products have no bullets.

Source subset

The data is the evaluation set of the master's dissertation cited below:

• ESCI, us locale, "Grocery & Gourmet Food" category (from the ESCI-S metadata).
• Products with descriptions of 1,500 to 2,000 characters: 3,121 products.
• Queries with 2 to 20 Exact, Substitute or Complement products in that set: 1,061 queries.
• Origin in ESCI: 814 queries from the train split and 247 from the test split. All of them are published in a single test split, because they were used only for evaluation.

Generation of the enriched queries

One request to OpenAI o3 per query, with this prompt:

Query: {query}

Rewrite the query as a natural and concise question that a user might ask a conversational recommender system to search for products in an e-commerce platform.
Ensure the question is short, clear, relevant, and conversational.
Respond only with the reformulated question, without adding any extra text or explanations.

The answers were kept verbatim, with no manual editing:

• Average length: 3.1 words (original) and 6.6 words (enriched).
• Five rewrites are imperatives ("Show me cocoa powder.") instead of questions.
• 15 queries share 6 rewrites, because the model normalizes spelling variants: "stroop waffle" and "stroopwafel" both become "Do you have stroopwafels?". Each query keeps its own labels.
• 185 rewrites contain typographic Unicode characters, mostly the non-breaking hyphen (U+2011), the right single quotation mark (U+2019) and the no-break space (U+00A0). Normalize them if your pipeline expects ASCII.

Usage

from datasets import load_dataset

repo = "daniloxvalle/esci-enriched-queries"
queries = load_dataset(repo, "queries", split="test")
qrels = load_dataset(repo, "qrels", split="test")
corpus = load_dataset(repo, "corpus", split="corpus")

# Text indexed for each product in the dissertation: title, description and bullets
docs = {p["product_id"]: " ".join([p["title"], p["description"], *p["bullets"]]) for p in corpus}

# Graded relevance for nDCG: {query_id: {product_id: gain}}
relevance = {}
for r in qrels:
    relevance.setdefault(r["query_id"], {})[r["product_id"]] = r["gain"]

Search docs with enriched_query (or with the original query, for comparison) and score the ranking against relevance. The query_id and product_id columns also join with the full ESCI and ESCI-S data (images, prices, other locales and labels).

Intended use

• Evaluating retrieval and RAG pipelines for product search with natural-language queries.
• Comparing keyword queries with conversational queries on the same relevance labels.

Limitations

• The rewrites were generated by an LLM; they are not real customer messages.
• Single-turn questions, with no conversation history.
• One product category (Grocery & Gourmet Food) and English only.
• A small catalog (3,121 products), much smaller than the full ESCI-S.

License

Apache 2.0, the same license as Amazon ESCI and ESCI-S.

Citation

@mastersthesis{valle2026design,
  title   = {Design and Evaluation of a {RAG}-Based Approach for Product Recommendation in E-Commerce},
  author  = {Valle, Danilo Xavier},
  school  = {Instituto Federal do Esp{\'i}rito Santo (IFES)},
  address = {Serra, ES, Brazil},
  type    = {Master's dissertation},
  year    = {2026}
}

@inproceedings{valle2025enhancing,
  title     = {Enhancing E-Commerce with a {RAG}-Powered Conversational Recommender System},
  author    = {Valle, Danilo Xavier and Oliveira, Hil{\'a}rio Tomaz Alves de},
  booktitle = {Anais do XXII Encontro Nacional de Intelig{\^e}ncia Artificial e Computacional (ENIAC 2025)},
  publisher = {Sociedade Brasileira de Computa{\c{c}}{\~a}o (SBC)},
  address   = {Fortaleza, CE, Brazil},
  pages     = {129--140},
  year      = {2025},
  doi       = {10.5753/eniac.2025.11835}
}

@misc{reddy2022shopping,
  title        = {Shopping Queries Dataset: A Large-Scale {ESCI} Benchmark for Improving Product Search},
  author       = {Reddy, Chandan K. and M{\`a}rquez, Llu{\'i}s and Valero, Fran and Rao, Nikhil and
                  Zaragoza, Hugo and Bandyopadhyay, Sambaran and Biswas, Arnab and Xing, Anlu and
                  Subbian, Karthik},
  year         = {2022},
  howpublished = {arXiv preprint arXiv:2206.06588}
}