OIRA course catalog assistant
A citation-based retrieval product for navigating Bucknell's academic catalog.
Status: University preview, with production deployment planned for Fall 2026.
Context
Course and program information lives across long catalog documents. Students and advisors need direct answers, but those answers must remain grounded in official university sources.
My role
Built with a three-person team for Bucknell's Office of Institutional Research and Analytics, supervised by a full-time staff member. I contributed to document retrieval, academic planning, conversation memory, and the chat interface, across a Next.js client and FastAPI services.
System
Source PDFs and Google Docs are ingested, split, embedded, and stored in ChromaDB. A query moves through the FastAPI backend to the retrieval layer, which selects relevant passages for OpenAI. The generated response returns with citations to the source material.
SQLite and UUID-based sessions preserve chat history. Clerk protects user-specific history and feedback. Response ratings and fallback instrumentation create a path for evaluating retrieval quality rather than treating generation as a black box.
Query classification, conversation summarization, and user-context extraction improve retrieval and response quality across multi-turn conversations.
Engineering decisions
- Ground every answer in retrieved university material and expose citations in the interface.
- Keep retrieval and generation behind a FastAPI boundary so the client remains independent of model orchestration.
- Persist session and feedback data separately from the vector store.
- Include OCR as an optional path for extracting information from schedule images.
Proposed backend refactor
In open PR #3, I reorganized the backend into typed FastAPI routes, services, repositories, and dependency injection. The proposal retains the existing chat, citation, streaming, session, feedback, and schedule interfaces. This work is under review and has not been merged.
Demo
Stack
Python, FastAPI, LangChain, ChromaDB, OpenAI API, SQLAlchemy, SQLite, Next.js, React, Tailwind CSS, Clerk, and Jest.