Enterprise RAG Chatbot — Job Finder Assistant

Apr – May 2026 · Associated with Miami University Farmer School of Business (ISA-632)
DatabricksApache SparkLangChain Delta LakeMLflowVector Search StreamlitLlama 4 Maverick

Built and publicly deployed an enterprise-grade Retrieval-Augmented Generation (RAG) chatbot for searching job listings via natural language queries. Built with a three-person team (Thenmozhi Boopathy, Seth Grace, and Luke Johnson) as part of the ISA-632 course at Miami University's Farmer School of Business, and deployed publicly on Streamlit Community Cloud.

Key features

A dual knowledge base combines 37 curated job PDFs (Deloitte, JPMorgan Chase, Amazon) with 2,530 scraped Data Engineer roles. Users search with natural language — e.g. “Show me remote managerial data analyst roles” — and get verified results; when nothing matches, the system explicitly explains what's available instead. A feedback loop collects thumbs up/down ratings with optional tags and comments (89% positive across 18 responses to date).

Architecture

Data ingestion parses PDFs via ai_parse_document() and transforms CSV data into a Delta table. Databricks Vector Search indexes the knowledge base, and a LangChain agent using the Llama 4 Maverick LLM with a VectorSearchRetrieverTool handles retrieval and reasoning. The model is served via a Databricks Model Serving endpoint, with a Streamlit UI that preserves chat history and collects user feedback, tracked with mlflow.genai.evaluate() for performance benchmarking.

Known trade-offs

Adding the CSV dataset improved coverage for entry-level roles but introduced retrieval noise, showing up as a drop in precision metrics during evaluation. The identified fix is hybrid search — combining semantic similarity with structured metadata filters on seniority, location, and skills.