Contribution · Application — E-commerce & Retail

AI Product Recommendations for E-commerce

Product recommendations are the highest-ROI surface on most e-commerce sites. Modern stacks pair dense embedding retrieval over the SKU catalog with LLM reranking for session-aware relevance, then fuse with business rules (stock, margin, brand pinning). The challenge is not the model — it's catalog quality, cold-start behavior, and avoiding hallucinated products that do not exist.

Application facts

Domain
E-commerce & Retail
Subdomain
Personalization
Example stack
OpenAI text-embedding-3-large or Cohere Embed v4 for catalog vectors · Pinecone or Weaviate for ANN retrieval over millions of SKUs · Claude Haiku 4.7 for session-aware reranking · Feature store (Feast) for user + SKU features · Algolia or Shopify Hydrogen for serving layer

Data & infrastructure needs

  • Structured SKU catalog with titles, descriptions, attributes, images
  • Clickstream and purchase history (session and user-level)
  • Inventory and margin data for business-rule fusion
  • Consent flags per DPDPA / GDPR for personalization

Risks & considerations

  • Hallucinated SKUs — model suggesting products that do not exist
  • Cold-start for new users or freshly listed items
  • DPDPA / GDPR compliance for profiling and behavioral targeting
  • Filter bubbles reducing catalog discovery and long-tail revenue

Frequently asked questions

Which model is best for e-commerce recommendations?

In April 2026, a Haiku-class model handles reranking at web latency for under a cent per session. Use Cohere Embed v4 or OpenAI text-embedding-3-large for the retrieval layer. The model matters less than catalog freshness and eval rigor.

Is LLM-based recommendation safe for retail?

Yes when grounded in the live catalog via constrained retrieval. Never let the LLM generate free-form product names — always return SKU IDs from a verified retrieval set and enforce stock and margin rules server-side.

Regulatory considerations for recommendation systems in India?

DPDPA 2023 requires explicit consent for profiling, purpose limitation, and the right to erasure. The Consumer Protection (E-commerce) Rules 2020 prohibit deceptive personalization. EU AI Act treats high-reach recommenders under transparency obligations.

Sources

  1. DPDPA 2023 — full text — accessed 2026-04-20
  2. Consumer Protection (E-commerce) Rules 2020 — accessed 2026-04-20

Related

  • LLM-Informed Dynamic Pricing for E-commerce — Dynamic pricing uses elasticity models and competitor signals to set SKU prices in near-real time — with LLMs adding narrative reasoning over promotions, inventory, and regulatory limits.
  • Semantic Product Search for E-commerce — Semantic search replaces brittle keyword lookup with embedding retrieval plus LLM query understanding — fixing zero-result pages, typos, and natural-language intent like 'gift for my father who likes cricket'.
  • AI Legal Case Law Research Assistant — AI legal research tools ground LLMs in curated case-law corpora (Westlaw, Manupatra, SCC Online) to produce cited, jurisdictionally-correct answers — avoiding the fabricated-citation disasters of ungrounded generative search.
  • All Applications — Browse the complete applications library
  • VSET — Programmes offered — B.Tech programs (CSE, AI&ML, AI&DS, Electronics, IIoT)