VineAI: Wine Pairing LLM

RAG Pipeline Architecture

The core innovation is a Retrieval-Augmented Generation pipeline that combines vector similarity search with LLM reasoning:

# User Input Processing
user_query = "I'm eating grilled salmon, budget $40"
embedding = embed(user_query)  # Convert to vector

# Vector Search
similar_wines = pinecone.query(
    vector=embedding,
    top_k=5,
    filter={"price": {"$lte": 40}}
)

# LLM Reasoning
recommendation = gemini.generate(
    prompt=f"""
    Customer: {user_query}
    Matching wines: {similar_wines}
    
    Provide personalized recommendation with:
    1. Best match wine with reasoning
    2. Tasting notes
    3. Why it pairs with salmon
    """,
    temperature=0.7
)
                  

Performance Optimization

• Caching layer: Redis for frequently requested pairings (90% cache hit rate)
• Embedding pre-computation: Wine profiles vectorized nightly
• Batch processing: Multi-table recommendations in single API call
• Edge deployment: Vercel Edge Functions for <100ms latency