sadrmanesh/llm-search-eval-electronics

llm-search-eval electronics example

A small bilingual (English and Persian) product search evaluation set from the llm-search-eval project. It has queries, a product catalog, graded gold labels, and the ranked results of three search systems, so you can try every part of an offline search evaluation end to end: metrics, LLM judges that fill label gaps, judge agreement, confidence intervals, pairwise comparison, and CI quality gates.

It is a worked example and a regression fixture, not a benchmark. Ten queries are far too few to rank real systems.

Try it in the browser: llm-search-eval demo Space · Docs: https://sadrmanesh.github.io/llm-search-eval/ · Package: llm-search-eval on PyPI

Contents

Config	Rows	Fields
queries	10	query_id, text, lang (en, fa, or empty)
products	22	product_id, title, brand, category, price, description
qrels	30	query_id, product_id, relevance (graded 0 to 3)
runs	107	system (bm25, semantic, hybrid), query_id, product_id, rank

Notes on the design:

• Five queries are English and five are Persian. q09 has no lang and uses Arabic letter forms (ي and ك), to exercise language detection and normalization.
• The catalog includes Persian listings and kitchen distractors.
• The gold labels are deliberately incomplete, so systems surface unlabeled products and a judge has gaps to fill.
• bm25 is a lexical baseline that is weak on Persian and returns nothing for q09. semantic is an embedding retriever and hybrid combines the two.

The original/ folder has the same data in the input formats the library reads directly: queries.jsonl, products.jsonl, qrels.csv, and one run per system as JSONL (runs/bm25.jsonl), CSV with scores (runs/semantic.csv), and a TREC run file (runs/hybrid.trec). The sample-report/ folder is the output of the command below: report.md, report.json, the self-contained report.html dashboard, cost.json, and judgments.jsonl.

Usage

pip install llm-search-eval
hf download sadrmanesh/llm-search-eval-electronics --repo-type dataset --local-dir electronics

llm-search-eval run \
  -q electronics/original/queries.jsonl \
  -p electronics/original/products.jsonl \
  --qrels electronics/original/qrels.csv \
  -s bm25=electronics/original/runs/bm25.jsonl \
  -s semantic=electronics/original/runs/semantic.csv \
  -s hybrid=electronics/original/runs/hybrid.trec \
  --judge mock --calibrate --pairwise-judge mock \
  -o reports

--judge mock is a deterministic offline judge that grades by query term coverage, which is useful for trying the pipeline. Use --judge openai=MODEL (or any OpenAI-compatible endpoint) for meaningful labels, and add --max-cost to cap the spend.

With the datasets library:

from datasets import load_dataset

queries = load_dataset("sadrmanesh/llm-search-eval-electronics", "queries", split="test")
runs = load_dataset("sadrmanesh/llm-search-eval-electronics", "runs", split="test")

License and citation

MIT, like the library. Product names are used only as realistic test data.

@software{sadrmanesh_llm_search_eval,
  author = {Sadrmanesh, Naser},
  title = {llm-search-eval: product search evaluation with IR metrics and LLM judges},
  url = {https://github.com/sadrmanesh/llm-search-eval},
  year = {2026}
}