Technology: RAG System

Components

• Document table from businesses_ranked.csv (fallback to clean)
• Embeddings with all-MiniLM-L6-v2
• FAISS cosine index (IndexFlatIP with L2-normalized vectors)
• Retriever combining tabular and semantic signals

Retrieval Flow

1. Parse intent (name/cuisine/rating/city/statistics)
2. Apply layered tabular retrieval (exact/fuzzy/category/city)
3. For semantic needs, query FAISS index
4. Merge, deduplicate, and score candidates
5. Build prompt context and optional passages

Prompt Rules

• Use only provided candidates and/or numbered passages
• Never invent restaurant names or facts
• Prefer higher review_count for authority
• Cite passages [1],[2] when used

Why This Works

• Tabular retrieval is precise for structured filters
• FAISS augments recall for natural language
• Strict prompt keeps LLM grounded → <2% hallucination