What is the best text embedding model for ecommerce product search (short, noisy user queries)?

[sumitrajrajsumit]

I am integrating a vector-based semantic search system into a large ecommerce platform's product search, and I want to select the right text embedding model.

Use Case
User queries are often:

• Very short (1–4 words)
• Ungrammatical
• Misspelled
• Contain specifications or abbreviations (e.g., “m12 nut”, “2hp pump”, “ss tank 1000l”)
• Contain domain-specific technical terms

Each product has:

• Title
• Attribute fields (e.g., Material=SS, Voltage=220V)
• Description text

I need embeddings that capture semantic meaning across these fields and match them with noisy, spec-heavy queries.

Constraints / Setup

• English-only
• Running on GPU (model size not a constraint)
• Throughput: ~100 queries per second
• Retrieval backend not yet decided but most likely Vespa
• Fine-tuning will come later — I first need a strong base embedding model

Questions

• Which open-source embedding models work best out of the box for ecommerce/product search?
• Are there any models that are trained or tuned specifically for ecommerce data?
• Should I embed (title + attributes + description) concatenated as a single document, or embed fields separately and combine?

Example queries

• “2hp motor pump”
• “ss nut m12”
• “isi water tank 1000l”
• “sewing macine” (misspelled)

Any guidance or practical experience with embedding models for ecommerce search would be appreciated.

[neyronius]

We have positive experience using snowflake-arctic-embed-m-v2.0 for e-commerce, but Marqo-Ecommerce-L would be a better match for your requirements.

[akkupratap323]

I would avoid choosing a single model before testing the retrieval design, because these queries contain two different signals:

• semantic intent: motor pump vs. water pump
• exact lexical/attribute constraints: 2hp, M12, SS, 1000L, standards such as ISI

A dense-only system is likely to blur the second category. My starting baseline would be:

1. Hybrid retrieval: a dense embedding plus BM25/sparse retrieval, fused with RRF. Exact model numbers, units, material codes, SKUs, and standards then survive even when the dense model treats them as weak tokens.
2. Canonical product text, for example title: ... | material: stainless steel (SS) | power: 2 hp | .... Normalize units and aliases, but retain the original tokens too. Keep truly filterable fields (brand, category, availability, numeric ranges) as structured payload/filter fields rather than asking an embedding to enforce them.
3. Separate representations only when you need separate weighting. If title and attributes should dominate description, use named vectors/fields and fuse them. Blind concatenation lets a long description dilute a short, high-value title; Qdrant's multi-representation guide describes this exact failure mode.

For an English, GPU-backed benchmark, I would test at least:

• Marqo/marqo-ecommerce-embeddings-B and -L as domain-specific candidates (the B model is a useful latency baseline; L is the quality candidate),
• Snowflake/snowflake-arctic-embed-m-v2.0 as a strong efficient text-retrieval baseline,
• your current model, so the experiment measures actual uplift rather than leaderboard position.

Marqo publishes ecommerce evaluation sets (AmazonProducts and GoogleShopping) and reports strong ecommerce results, but those results do not replace an evaluation on your catalogue and query distribution.

Build a small judged set from real queries/clicks and include hard negatives from the same category with one wrong specification (2 hp vs 1 hp, M12 vs M10, 1000 L vs 500 L). Compare Recall@20/100 and nDCG or MRR, then measure end-to-end p95 latency and GPU throughput at your 100 QPS target. Fine-tune only after inspecting which failure class remains; hard-negative training on specification mismatches is usually more valuable than generic extra pairs.

For typo handling, normalize/expand the query before retrieval (units, common abbreviations, spelling), but search with both the original and normalized form so a correction cannot erase an exact product code.

Qdrant references for the retrieval shape:

• Hybrid dense+sparse search: https://qdrant.tech/documentation/search/text-search/hybrid-search/
• Multiple title/field/chunk representations: https://qdrant.tech/documentation/tutorials-search-engineering/multi-representation-search/
• Ecommerce hybrid example with filters: https://qdrant.tech/e-commerce/

Model cards/evaluation data:

• https://huggingface.co/Marqo/marqo-ecommerce-embeddings-B
• https://huggingface.co/Snowflake/snowflake-arctic-embed-m-v2.0