Enterprises face a growing challenge: business-critical documents stored in SQL databases weren't designed for semantic search. A new hybrid architecture demonstrates how to implement Agentic RAG systems directly on existing SQL tables without schema changes, data migration, or performance trade-offs.

Traditional relational databases store proposals, reports, and contracts in TEXT columns optimized for keyword matching. Modern demands require deep semantic understanding and contextual insights. The solution combines SQL computation with vector-based semantic search using a ReAct agent that intelligently routes queries. The system uses FAISS for vector embeddings and gemini-2.5-flash as the LLM, though the design works with any vector database or language model.

The architecture employs two specialized tools: a SQL tool for computations and aggregations, and a vector tool for semantic content search. The ReAct agent determines which tool to invoke based on query type—SQL for calculations and ranges, vector for topic searches, and hybrid approaches when both structured filtering and semantic understanding are needed. Results demonstrate the system's effectiveness across three query types: computational queries like "articles by year and category," semantic searches like "articles about children," and hybrid queries combining metadata filters with semantic matching.