ManmohanBuildsProducts/auto-parts-search-benchmark

Auto-Parts-Search Benchmark

A public, reproducible retrieval benchmark for Indian automotive parts search — Hindi / Hinglish / English / misspellings / brand-as-generic / symptom / part-number query types. Built to let anyone independently verify claims about domain-tuned vs commercial embedding models on this task.

Companion to the model ManmohanBuildsProducts/auto-parts-search-v3 and the full eval writeup in docs/EVAL_REPORT.md.

What's inside

Split	File	Queries	Use
dev	benchmark_dev.json	149	Tuning + reporting
test (SEALED)	benchmark_test.json	46	Do not tune on this. Reserved for follow-up publication.

Each query:

{
  "query": "engine garam ho raha",
  "query_type": "symptom",
  "expected_parts": ["radiator", "thermostat", "coolant"],
  "expected_categories": [],
  "expected_vehicles": [],
  "difficulty": "easy",
  "notes": ""
}

Query types (149 dev queries):

Type	n	Example
exact_english	27	"brake pad for Swift"
misspelled	23	"break pad"
hindi_hinglish	27	"गाड़ी का brake" or "gaadi ki batti"
symptom	27	"engine garam ho raha"
brand_as_generic	22	"Mobil for Swift", "Amaron for Innova"
part_number	23	"6U7853952"

Graded relevance labels

round1_graded.jsonl and round2_graded.jsonl contain LLM-judged relevance grades for each (query, candidate) pair:

• 2 = RELEVANT (direct match)
• 1 = MARGINAL (related but not direct)
• 0 = IRRELEVANT

Judge: DeepSeek V3 (chat), temperature 0. Prior calibration showed 96% query-level agreement with Claude Opus on this task.

Pool construction: for each query, the top-20 hits from every evaluated model are unioned into a joint pool (median ~60 docs), then judged once. This eliminates per-model bias — every model is scored against the same judged pool.

Corpora

• Round 1 — KG corpus (2,121 docs): curated auto-parts knowledge graph entries (names + aliases + parent systems).
• Round 2 — Production corpus (26,835 docs): the full index behind our production /search endpoint. 884 KG + 25,951 scraped catalog products (SparesHub, Eauto, Bikespares).

round2_corpus.json contains the (id, text, metadata) triples for the 26,835 production corpus.

Results summary (round 2, production corpus)

Model	nDCG@10	R@5	P@1	MAP@10
openai/text-embedding-3-large	0.468	0.290	0.550	0.477
v3+BM25 hybrid (tuned, 3-fold CV)	0.424	0.257	0.523	0.458
v3 (embedding only)	0.411	0.240	0.503	0.439
cohere/embed-multilingual-v3.0	0.332	0.218	0.456	0.379
intfloat/multilingual-e5-large	0.309	0.221	0.450	0.360
BAAI/bge-m3	0.307	0.194	0.403	0.328

Full results + per-category breakdowns: see round2_scores.json and EVAL_REPORT.md.

How to use

from datasets import load_dataset
ds = load_dataset("ManmohanBuildsProducts/auto-parts-search-benchmark", data_files="benchmark_dev.json")
print(ds["train"][0])

Or just:

import json
queries = json.load(open("benchmark_dev.json"))

Reproducibility

Full pipeline at github.com/ManmohanBuildsProducts/auto-parts-search. Bench scripts:

• scripts/bench_external.py — round 1 (KG only)
• scripts/bench_production.py — round 2 (production corpus)
• scripts/tune_hybrid.py + scripts/tune_hybrid_cv.py — fusion-weight tuning

Cost to reproduce full round 2: ~$0.20 OpenAI embeddings + ~$0.30 DeepSeek judge. Cohere/Jina fit within free tiers.

Limitations

1. n = 149 dev queries. Overall nDCG@10 MoE ≈ ±3–5% at 95% CI; per-category (n=22–27) MoE ≈ ±10%. Small per-category deltas are noise.
2. LLM judge, not human. DeepSeek V3 at temperature 0, calibrated against Claude Opus at 96% agreement.
3. Indic auto-parts-specific. Does not generalize to other e-commerce domains.
4. The hybrid's BM25 uses a custom Indic tokenizer (Roman↔Devanagari expansion) — third parties running vanilla BM25 will get different hybrid numbers.

Citation

@misc{auto_parts_search_bench_2026,
  author    = {Khurana, Manmohan},
  title     = {Auto-Parts-Search: Indian Automotive Retrieval Benchmark + v3 Model Card},
  year      = 2026,
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/ManmohanBuildsProducts/auto-parts-search-benchmark},
}

License

• This benchmark dataset: CC-BY-4.0 (attribution required; any use incl. commercial).
• Companion v3 model: Apache-2.0 (inherits from BGE-m3 MIT, upgraded to Apache-2.0 for patent grant).

Contact

Manmohan Khurana — manmohanbuildsproducts [at] gmail.com