Abstract
Hybrid information retrieval systems are needed more and more in enterprise environments where both structured data and unstructured data reside. Retrieval-Augmented Generation (RAG) systems are suited for descriptive, definition-oriented, and semantic queries but are not robust enough for analytical or aggregation-based queries. Textto-SQL methods are good at dealing with structured analytical tasks but are unable to capture the context-rich semantics of unstructured data. In this paper, we introduce a Hybrid Agentic AI Framework that unites RAG and Text-to-SQL techniques under a multi-agent orchestration model. Structured information is kept in PostgreSQL, where each row is also converted to textual abstract (key-value form) and embedded for semantic search. Unstructured document data (PDFs, Word, PPTs) is tokenized into chunks and embedded for similarity retrieval. A group of specialized agents work in concert to interpret user queries: (i) a Temporal Agent derives time constraints, (ii) an Entity Agent detects important entities and uses them as filters, (iii) a Decision Agent chooses between RAG and Text-to-SQL paths, and (iv) an SQL Generation Agent generates optimized queries if the SQL path is selected. This combined approach helps precise query resolution: count based and aggregation based queries are processed by SQL, while contextual and descriptive queries are processed by RAG. After temporal and entity filtering, results are synthesized by a Large Language Model (LLM) into natural language responses. This method proves the effectiveness of agentic orchestration in enhancing both precision and semantic depth over heterogeneous data sources.