Mercado Libre, Latin America's largest e-commerce platform, addressed the challenge of handling complex search queries by implementing vector embeddings and Google's Vector Search database. Their traditional word-matching search system struggled with contextual queries, leading to irrelevant results. The new system significantly improved search quality for complex queries, which constitute about half of all search traffic, resulting in increased click-through and conversion rates.
This case study comes from a Google Cloud presentation featuring two Latin American companies: Mercado Libre and Grupo Boticário. The presentation, hosted by Audrey Chang (Director at Google Cloud’s Global Strategic Initiatives team), showcases how these companies are exploring and implementing generative AI technologies to improve customer experiences and create operational efficiencies. This summary focuses primarily on the Mercado Libre case study as presented by Nicholas Presta, Senior Engineering Manager of Machine Learning.
Mercado Libre is described as the largest online commerce and payments platform in Latin America, serving approximately 144 million unique active users. The company’s vision centers on democratizing commerce and finance across the continent through a trusted, agile, and people-centric platform. The scale of this operation presents significant challenges in contextualizing and serving user experiences in real-time.
The core problem that Mercado Libre faced was the inadequacy of their traditional search engine approach, which relied heavily on word matching. Nicholas Presta explained this limitation through concrete examples that illustrate the frustration users experienced.
When a user searched for “a computer that allows me to play Fortnite and make video edition,” the traditional system would surface irrelevant results like computer books or video games rather than actual computers with gaming capabilities. Similarly, when searching for “I want a gift for my daughter who loves soccer and is a fan of Lionel Messi,” the system would suggest items like a heart pendant with the word “daughter” engraved on it rather than soccer-related items or Messi merchandise.
The traditional word matching approach works by finding all items containing the specific words in the query (such as “Samsung” and “TV”) within item titles or attributes, then fetching, ranking, and displaying these results. While this approach works adequately for simple, direct queries, it fundamentally fails when users express complex intents, use natural language, or have nuanced requirements that cannot be captured through exact keyword matching.
This limitation is particularly significant given the scale of Mercado Libre’s operations. The presentation notes that most successful purchases on the platform begin with a search query, and approximately half of all distinct queries made by users are classified as “complex queries.” This means that improving search quality for complex queries has direct and substantial business impact on click rates and conversion rates.
To address these limitations, Mercado Libre implemented a semantic search approach using vector embeddings combined with Google Cloud’s Vector Search database. The technical architecture works as follows:
First, the system generates a vector embedding for each item in the catalog. These embeddings capture the semantic meaning and characteristics of products in a high-dimensional vector space, allowing items with similar meanings or attributes to be positioned closer together mathematically, even if they don’t share the same keywords.
All these item embeddings are then inserted into Google’s Vector Search database, which is optimized for efficient similarity searches across large-scale vector datasets. This infrastructure choice is critical given the massive catalog that Mercado Libre likely maintains across Latin America.
When a user submits a query, the system computes an embedding for that query text. This query embedding captures the semantic intent of what the user is looking for, including implicit meanings that might not be captured by simple keyword extraction. The system then asks Vector Search to find the closest matching items to that query embedding in the vector space.
This approach enables the search engine to understand that a query about “a gift for my daughter who loves soccer and is a fan of Lionel Messi” should return Messi jerseys, soccer equipment, and related merchandise, even if these items don’t explicitly contain the word “daughter” or “gift” in their descriptions. The semantic understanding bridges the gap between user intent and product catalog in ways that keyword matching fundamentally cannot.
While the presentation does not go into extensive detail about the operational aspects of running this system in production, several LLMOps considerations can be inferred from the discussion:
Scale and Performance: Serving 144 million unique active users requires the embedding generation and vector search infrastructure to operate at massive scale with low latency. Search is a real-time, synchronous operation where users expect results within milliseconds. This implies significant investment in infrastructure for both the embedding models and the vector database.
Continuous Learning and Evolution: Audrey Chang’s comments about how “longtail queries will change over time” and how the system will need to understand queries that were previously considered longtail suggest that the embedding models need to be periodically retrained or fine-tuned as user behavior patterns evolve. The concept of queries becoming “no longer longtail” as the system learns them implies an ongoing MLOps cycle of model improvement.
Two-Sided Marketplace Benefits: An interesting aspect highlighted in the discussion is that the improved search quality benefits both buyers and sellers. Buyers find what they need more efficiently, while sellers connect more effectively with buyer needs. This dual benefit creates strong business justification for continued investment in the generative AI infrastructure.
Integration with Existing Systems: The presentation suggests that the vector search approach augments rather than replaces the existing search infrastructure. For straightforward queries where word matching works well, that approach likely continues to function. The vector-based approach appears to be particularly targeted at improving results for complex queries where traditional methods fail.
The presentation also features Marcus Bittencourt from Grupo Boticário, a major Brazilian beauty and cosmetics company. Their approach differs from Mercado Libre’s search use case, focusing instead on internal efficiency and customer service applications.
Grupo Boticário developed what they call a “Gen AI platform” with multiple LLMs, designed around their concept of “Love Tech” – using technology to enhance human capabilities rather than replace people. Their implementations include chatbots for sales representatives that answer simple questions, freeing up staff for higher-value activities, and first-response automation for franchising inquiries that reportedly achieves 70% accuracy in action classification.
They’ve also developed what they call “gen self-learning services” – a platform that enables non-technical teams to work with generative AI and create context-aware applications without requiring deep technical expertise. This democratization of AI capabilities represents an interesting approach to scaling AI adoption within an organization.
It’s important to note that this presentation is a Google Cloud showcase, and both companies are presented as successful customers of Google Cloud technologies. As such, the results are likely presented in the most favorable light possible.
Some aspects that would benefit from more detail include:
The 70% accuracy figure mentioned for Grupo Boticário’s first-response automation is a useful concrete metric, though it would be helpful to understand what the remaining 30% of cases look like and how they are handled.
The Mercado Libre case study represents a classic and compelling application of vector embeddings and semantic search in e-commerce. The fundamental insight that half of user queries are “complex” and poorly served by keyword matching creates strong justification for investment in semantic search technologies. The use of Google Cloud’s Vector Search as the underlying infrastructure demonstrates how managed services can enable large-scale deployment of embedding-based systems. While the presentation lacks detailed quantitative results, the described improvements in search quality for complex queries represent meaningful progress in serving users with natural language queries in e-commerce contexts.
Snorkel developed a specialized benchmark dataset for evaluating AI agents in insurance underwriting, leveraging their expert network of Chartered Property and Casualty Underwriters (CPCUs). The benchmark simulates an AI copilot that assists junior underwriters by reasoning over proprietary knowledge, using multiple tools including databases and underwriting guidelines, and engaging in multi-turn conversations. The evaluation revealed significant performance variations across frontier models (single digits to ~80% accuracy), with notable error modes including tool use failures (36% of conversations) and hallucinations from pretrained domain knowledge, particularly from OpenAI models which hallucinated non-existent insurance products 15-45% of the time.
DoorDash faced challenges in scaling personalization and maintaining product catalogs as they expanded beyond restaurants into new verticals like grocery, retail, and convenience stores, dealing with millions of SKUs and cold-start scenarios for new customers and products. They implemented a layered approach combining traditional machine learning with fine-tuned LLMs, RAG systems, and LLM agents to automate product knowledge graph construction, enable contextual personalization, and provide recommendations even without historical user interaction data. The solution resulted in faster, more cost-effective catalog processing, improved personalization for cold-start scenarios, and the foundation for future agentic shopping experiences that can adapt to real-time contexts like emergency situations.
eBay developed Mercury, an internal agentic framework designed to scale LLM-powered recommendation experiences across its massive marketplace of over two billion active listings. The platform addresses the challenge of transforming vast amounts of unstructured data into personalized product recommendations by integrating Retrieval-Augmented Generation (RAG) with a custom Listing Matching Engine that bridges the gap between LLM-generated text outputs and eBay's dynamic inventory. Mercury enables rapid development through reusable, plug-and-play components following object-oriented design principles, while its near-real-time distributed queue-based execution platform handles cost and latency requirements at industrial scale. The system combines multiple retrieval mechanisms, semantic search using embedding models, anomaly detection, and personalized ranking to deliver contextually relevant shopping experiences to hundreds of millions of users.
