Glowe represents a sophisticated application of LLMOps principles in the beauty and skincare domain, demonstrating how specialized AI agents can provide expert-level recommendations in complex, nuanced fields. The case study showcases Weaviate's approach to building a Korean skincare recommendation system that goes beyond generic product matching to understand actual skincare science and ingredient interactions. ## Problem and Business Context The skincare industry generates thousands of products annually with millions of reviews, yet most recommendation systems fail to capture the nuanced relationships between ingredients, user experiences, and skincare outcomes. Traditional systems often rely on generic product embeddings that dilute the importance of specific skincare effects, making it difficult to build effective, personalized routines. The challenge was particularly acute in Korean skincare, where complex multi-step routines require understanding ingredient compatibility, timing, and layering principles. ## Technical Architecture and LLMOps Implementation The solution employs a multi-layered LLMOps architecture centered around domain-specific agents and sophisticated embedding strategies. At its core, Glowe uses a dataset of 1,440 curated Korean skincare products and over 94,500 user reviews, all processed through an LLM pipeline using Gemma 3 12B to extract actual effects rather than relying on marketing descriptions. ### Dual Embedding Strategy A critical innovation in Glowe's approach is the implementation of a dual embedding strategy using Weaviate's named vectors capability. Rather than embedding entire product objects into single vectors, the system separates product representation into two distinct embeddings: The **Product Embedding** captures descriptive metadata including name, brand, category, and ingredients, primarily used for general browsing and product search. The **Effect Embedding** focuses specifically on skincare benefits extracted from user reviews, which is crucial for matching products to user skin profiles and building effective routines. ### TF-IDF Weighted Effect Embeddings The effect embedding employs a sophisticated TF-IDF weighting mechanism to understand not just what effects a product provides, but how strongly it's associated with specific outcomes. The system first processes reviews through LLM analysis to extract positive and negative effects, then clusters these into 250 standardized effects to handle aliases and semantic variations. The TF-IDF calculation considers both term frequency (how often an effect is mentioned for a specific product) and inverse document frequency (how rare the effect is across all products). This ensures that rare but meaningful effects aren't drowned out by common ones like basic hydration. The final effect vector is constructed by blending static effect vectors using these calculated weights, with optional description diversity mixing to avoid duplicates. ### Production LLM Pipeline The routine generation process demonstrates sophisticated prompt engineering and LLM orchestration. User information is collected through a structured questionnaire and formatted into a comprehensive profile including goals, conditions, skin type, sensitivities, and age. The system then uses category selection logic to determine which product types are needed for morning and evening routines. Product retrieval leverages Weaviate's vector search capabilities, running similarity searches against the effect vectors to find products matching specific skin needs. This is executed concurrently across all selected categories to create shortlists of top-matching products. The final assembly uses Gemini 2.5 Flash with detailed domain-specific prompting to select optimal products from each category while ensuring compatibility, respecting sensitivities, and avoiding ingredient conflicts. The LLM also generates metadata including ingredient notes and suggestions to help users understand their routines. ### Agentic Chat Interface with Elysia Glowe implements an advanced agentic chat interface powered by Elysia, Weaviate's open-source agentic platform. This represents a sophisticated approach to conversational AI that goes beyond simple query-response patterns. Elysia operates on decision-tree logic, enabling guided conversations based on user input and context, with built-in user management for conversation history tracking. The system demonstrates advanced tool integration capabilities through custom agents like the Stack Agent and Product Agent. A key component is the Weaviate Query Agent, which translates natural language queries into structured searches across Weaviate collections, handles complex multi-stage questions, and synthesizes responses with full transparency about data sources and query methods. ### LLMOps Production Considerations The case study reveals several important LLMOps practices in production deployment. The system implements self-healing and correction loops within the decision-making cycle, allowing the agent to identify misunderstandings and attempt recovery for more robust interactions. Context awareness is maintained across conversations, with proper user session management and history tracking. The architecture demonstrates proper separation of concerns with specialized agents handling different aspects of the recommendation process. The dual embedding approach shows sophisticated data engineering that optimizes for both general search and domain-specific matching requirements. ## Evaluation and Performance Insights While the case study doesn't provide specific performance metrics, it emphasizes the importance of grounding recommendations in actual user experiences rather than marketing claims. The TF-IDF weighting system ensures that products are ranked by the strength of their relevant effects, which is crucial for building effective skincare routines. The domain-specific approach appears to provide significant advantages over general-purpose recommendation systems by understanding skincare science, ingredient interactions, and routine timing principles. The agentic interface enables complex, contextual conversations that can handle follow-up questions, routine modifications, and alternative recommendations. ## Technical Innovations and LLMOps Best Practices Several technical innovations stand out in this implementation. The dual embedding strategy with TF-IDF weighting represents a sophisticated approach to handling domain-specific recommendation requirements. The use of named vectors in Weaviate allows for independent searchability of different aspects of products, which could be applicable to other complex recommendation domains. The integration of multiple LLMs for different purposes (Gemma 3 12B for review analysis, Gemini 2.5 Flash for routine generation) demonstrates proper model selection based on task requirements. The agentic framework with Elysia shows how complex conversational AI can be built with proper tool integration and decision-tree logic. The emphasis on domain knowledge agents rather than generic AI solutions reflects a mature understanding of how to deploy LLMs effectively in specialized fields. The system architecture supports both batch processing (for review analysis and embedding generation) and real-time interaction (for chat and routine generation). ## Scalability and Production Deployment The architecture appears designed for scalability with concurrent processing capabilities for product retrieval and proper separation between data processing and user-facing services. The use of Weaviate as the vector database provides enterprise-grade capabilities for handling large-scale embedding storage and similarity search operations. The modular design with specialized agents allows for independent scaling and updating of different system components. The open-source approach with Elysia suggests a commitment to reproducibility and community development, which is important for long-term maintainability in production LLMOps environments. ## Limitations and Considerations While the case study presents compelling results, it's important to note that it primarily represents a proof-of-concept and promotional content from Weaviate. The actual effectiveness of the recommendations in practice would require independent validation and user studies. The focus on Korean skincare, while providing domain specificity, may limit generalizability to other beauty markets with different product types and cultural practices. The system's reliance on user-generated review data for effect extraction could introduce biases based on review demographics and product accessibility. The LLM-based review analysis, while sophisticated, may not capture all nuanced effects or may introduce interpretation errors that could propagate through the recommendation system. Overall, Glowe demonstrates sophisticated LLMOps practices for building domain-specific AI applications, particularly in the areas of specialized embeddings, agentic architectures, and production-ready conversational AI systems.