Work

Enterprise RAG System for Structured & Unstructured Intelligence

RAG
LLM
LangChain
FAISS
Streamlit

A hybrid AI agent system merging natural language SQL queries over a 50K+ record inventory database with FAISS-powered semantic search across corporate policy documents — all through a single Streamlit dashboard.

Overview

Built a production-grade hybrid intelligence system that lets non-technical users interact with both structured inventory data and unstructured corporate policy documents using plain natural language. The system intelligently routes each query to the right data source — SQL for analytical questions, RAG for policy lookups — and handles hybrid queries that need both.

GitHub: saideepa05/Enterprise_inventory

What I Built

  • Hybrid query orchestration engine — a smart routing layer built with LangChain LCEL that detects query intent and dispatches to either the SQL module, the RAG pipeline, or both simultaneously for hybrid queries
  • SQL intelligence module — natural language interface over a 50,000+ record diamond inventory SQLite database, supporting complex analytical queries like “What is the average price of Ideal cut diamonds?”
  • Document intelligence (RAG pipeline) — FAISS vector store with semantic search over corporate policy documents (grading policies, return procedures, approval workflows), enabling queries like “What is the restocking fee for high-end returns?”
  • ETL & data pipeline — automated data ingestion from CSV, database initialization scripts, and document chunking/embedding pipeline from raw text to indexed vector store
  • Streamlit dashboard — professional UI with glassmorphic design, unifying both data sources in a single responsive interface accessible to non-technical stakeholders

Tech Stack

Python · LangChain (LCEL) · OpenAI API · FAISS · SQLite · Streamlit · Pandas · Embedding Models

Impact

Unified two fundamentally different data sources — a relational inventory database and unstructured policy documents — behind a single natural language interface. Non-technical users can now run queries that previously required a developer or data analyst, including hybrid questions that cross both structured and unstructured data in a single response.