Walmart uses LLMs to enrich 850 million product data points — a task that would have needed 100x the headcount manually

Background

Walmart's product catalog contains billions of data points. Inaccurate or missing attributes led to poor search results, disappointed customers, and operational inefficiencies for store and warehouse associates trying to locate items. Manual catalog maintenance was impractical at the scale Walmart operates.

What Was Implemented

• Large language models (third-party and proprietary) applied to product catalog attribute extraction
• Multi-agent pipeline: attribute extraction LLM + quality-check LLM trained on human-validated ground truth
• Multi-task fine-tuning and knowledge distillation to reduce inference cost while maintaining accuracy
• Model optimization via LoRA, quantization, gradient checkpointing, and Flash Attention
• Integration with store associate mobile tools for real-time product location
• LLM-generated predictions connected to Walmart's AI shopping assistant

Results

- More than 850 million pieces of product data created or improved via LLM - Process achieves what would have required 100 times the headcount if done manually, per executive statements - Store associates can now locate inventory quickly via mobile devices, replacing the prior "treasure hunt" - Customers using Walmart's AI shopping assistant receive improved product recommendations and follow-up answers

Lessons

• Multi-agent LLM architectures (extraction + quality check) outperform single-model approaches for catalog enrichment at retail scale
• Human-validated training data remains essential: ground truth labels from human specialists define the standard the QC model learns to match
• Model optimization techniques (LoRA, quantization) make large-scale catalog enrichment economically viable on constrained compute
• Catalog data quality has downstream effects across operations — search, fulfillment, inventory, and associate workflows all benefit