An AI-powered retail recommendation system with voice input, natural language processing, and intelligent product retrieval using Retrieval-Augmented Generation (RAG). The system helps customers find the right products through a conversational interface.
Reta-AI is a conversational retail assistant that enables product discovery using:
- Vector embeddings for semantic search over 2,000 products
- Structured intent parsing (11-field schema)
- Priority-based constraint filtering with intelligent relaxation
- Gemini-powered natural language response generation
- Voice input with transcription and fuzzy correction
- Store navigation using color-based location mapping
- Web interface built with Next.js
- Full dataset integration using all 18 columns
- Enhanced 11-field intent schema
- Priority-based 9-step constraint engine
- Precomputed alternate product recommendations
- Store location mapping using color zones
- Rich contextual responses using full dataset
The application is deployed and accessible at:
Web Link: https://retaai.vercel.app/
Note: The first response may take a few seconds due to cold start. Subsequent responses are faster.
- Semantic search using FAISS with normalized 384-dimensional embeddings
- Structured intent extraction (category, brand, size, budget, seasonal, etc.)
- Intelligent filtering using ordered constraint pipeline
- Context-aware recommendations with alternates
- Voice input with improved transcription accuracy
- Natural language responses with product and location details
- Real-time web-based chat interface
- Coverage of 2,000+ products
Frontend (Next.js)
- Chat interface with voice and text input
Backend (FastAPI)
- Query parsing
- FAISS vector retrieval
- Constraint filtering
- Recommendation engine
- Response generation using Gemini
Data Layer
- 18-column product dataset
- FAISS index for similarity search
- Metadata stored as DataFrame
Converts user queries into structured 11-field schema including category, brand, size, budget, and preferences.
Applies filters in priority order:
- Stock availability
- Category
- Brand (with relaxation if needed)
- Price
- Size
- Seasonal
- Source
- Exclusions
- FAISS-based vector search
- Normalized embeddings (384-dim)
- Deduplication of products
- Selects best product
- Uses precomputed alternate product IDs
- Adds store location hints
- Uses Gemini for natural language output
- Includes product details, alternatives, and location guidance
- Backend: Python (FastAPI)
- Retrieval: FAISS
- LLM: Gemini
- Frontend: Next.js
- Hosting: AWS EC2 (backend), Vercel (frontend)
- Retrieval accuracy improved using structured embeddings
- Faster filtering using priority-based constraints
- Reduced latency through simplified pipeline
- First request may experience cold start delay
Refer to the following for more details:
- PIPELINE_UPGRADE_SUMMARY.md
- INTEGRATION_GUIDE.md
- CUSTOMER_TEST_CASES.md