Be-Cause.shop : Recommandations de Mode Éthique Alimentées par l'IA

We deployed for Be-Cause.shop an innovative recommendation system combining RAG and LLM, enabling personalized suggestions. This optimized cloud solution has significantly improved user experience while maintaining controlled operational costs.

ABOUT THE CLIENT

Be-Cause.shop stands as a pioneer in ethical and responsible fashion. The platform brings together a committed community of sustainable brands and conscious consumers, creating a unique ecosystem in the online fashion sector.

CHALLENGE

Be-Cause.shop faces three major challenges in its mission to offer a personalized shopping experience:

Database Connections

Integration and synchronization with multiple product data sources require a robust architecture capable of handling different data formats and structures while maintaining their consistency.

Relevant AI Matching

Developing AI matching algorithms capable of effectively analyzing user requests and providing truly relevant recommendations is a major technical challenge. The AI must understand users' real preferences and adapt to their evolving needs.

Processing Speed

Performance is crucial: users must instantly receive results matching their expectations. This requires advanced optimization of response times, from preference analysis to recommended product display.

SOLUTION

Our approach revolves around three main components:

RAG Recommendation System

• Vector/embeddings calculation for each product

• Use of MongoAtlas for embeddings storage

• Intelligent product indexing for ultra-fast response time

• Advanced semantic search

LLM Validation

• Integration of GPT-4o Mini for cost optimization

• Verification of recommendation consistency and adjustment

• Generation of personalized explanations

Cloud Infrastructure

• Cloud deployment for optimal scalability

• Operating cost optimization

• Real-time performance monitoring

METHODOLOGY

The project was carried out with a tight team consisting of:

• 1 AI expert

• 1 full-stack developer

• 1 product owner

Development was conducted in two-week iterations, with systematic testing phases and gradual deployment to staging then production.

CONCRETE EXAMPLE

User Request:

"I'm looking for casual office wear that's both elegant and eco-responsible. I prefer neutral tones and have a budget of around €200."

AI Analysis:

1. RAG identifies key criteria:

   • Use: casual office wear
   • Style: elegant but casual
   • Values: eco-responsible
   • Color preference: neutral tones
   • Budget: €200

2. LLM validates consistency and generates personalized response:

"Here's a selection that matches your eco-responsible business casual style:

Proposed Outfit:

• Beige Tencel™ fluid shirt from EcoWear - €75 Made in France from biodegradable fibers

• Taupe organic cotton straight pants from GreenStyle - €89 Fair trade certified production

• Accessory: Natural cork belt from Cork Fashion - €35 Vegan leather alternative

Total: €199

These pieces are easily combinable with each other and your existing wardrobe. Tencel™ offers excellent comfort for the office while being environmentally friendly. Would you like to see other combinations in these tones?

RESULTS

Our recommendation system combining RAG and LLM pushes the boundaries of e-commerce personalization. By combining the power of semantic search with language models, we offer users a truly personalized shopping experience:

Technical Performance

• Fast response time thanks to cloud optimization
• Scalable system according to query volume
• Controlled operational costs thanks to GPT-4 Mini usage

User Experience

• More relevant recommendations based on specific requests
• Suggestions adapted to user requests
• Personalized responses with clear explanations

Business Impact

• Better product visibility
• Effective highlighting of brand commitments
• Facilitation of product discovery matching user criteria

PERSPECTIVES

The perspectives following the success of this first phase open the way to several promising developments: integration of predictive trend analysis, consideration of similar purchases, continuous improvement of the recommendation model, and enhanced personalization of explanations.