Vinny — AI Wine Concierge

The Problem

Wine recommendations are either shallow (filter by price/region) or require expensive sommelier expertise. Existing AI tools hallucinate wine names and tasting notes because they lack grounding in real inventory data.

Approach

Two-tier RAG pipeline combining pgvector semantic search with tsvector full-text search, fused via Reciprocal Rank Fusion, then reranked by Cohere — catching both conceptual queries ("bold Italian red") and exact lookups ("2019 Barolo").

Architecture

Next.js 16 App Router with Vercel AI SDK v6 orchestrates OpenAI GPT-4.1 tool calls against Supabase (pgvector 512-dim + tsvector hybrid search + RRF k=60 + Cohere reranking). Additional tools include Grapeminds API, Tavily web search, and an MCP server for extensible tool protocol.

Results

• 19 development phases complete
• 107K+ data sources integrated
• Hybrid search with measured precision improvements over vector-only
• Multi-tenant B2B SaaS architecture with Row-Level Security
• Live at vinny.chat

Lessons Learned

Hybrid search (vector + keyword + reranking) dramatically outperforms vector-only retrieval for domains where users mix conceptual and exact-match queries.