BrunoRAG

Live at brunorag.com

Abstract

BrunoRAG is a search service that allows users to query for courses at Brown University using natural language. Queries are embedded with OpenAI text-embedding-3-small and matched against ~131k course section vectors via FAISS. A pre-query LLM extracts structured filters (days, hours, ratings, instructor, etc.) from the input, and a post-query LLM generates a cited natural-language answer from the top results.

Data

Course data is collected from the Courses@Brown (CAB) API: ~131,000 section records spanning Summer 2016 through Spring 2026, representing ~10,000 unique offerings. Critical Review data from 2019–2025 is joined by CRN where available, providing course/professor ratings, workload hours, and class size.

A refresh pipeline (pipeline/refresh_data.py) discovers new CAB terms, fetches only changed data, preserves Critical Review fields, and regenerates stale embeddings automatically:

python3 pipeline/refresh_data.py --dry-run   # preview changes
python3 pipeline/refresh_data.py --apply      # fetch, embed, rebuild

Supported Queries

A pre-query model automatically extracts filters from natural language. Example queries:

• like math 1530 but more chill
• drama intensive MWF 4+ instructor rating
• WRIT classes in economics
• english classes <10 hours/wk MWF
• TuTh systems-heavy CS courses
• Cities and urban architecture practicum

Filterable fields: department, days of the week, semester/year, permission required, average/max weekly hours, course rating, professor rating, class size, instructor name, concentration programs.

Methodology

Each search request runs two LLM calls:

1. Pre-query (JSON mode) — extracts structured filters and produces a cleaned semantic query. Department filters use a two-stage FAISS search: primary department first, then fallback to broader results if top-K isn't filled.
2. RAG response — receives top-40 filtered and top-40 unfiltered FAISS results, generates a cited answer.

Course data is embedded as 256-dimensional vectors and indexed in a FAISS IndexFlatIP (cosine similarity) index loaded in-process at startup. An LRU cache (30min TTL, 100 entries) with in-flight deduplication serves repeated queries without re-running the pipeline.

UI

The frontend is a single-page React app (Vite + React 19). Features:

• Search bar with example query cards on the landing page
• "Only search Fall '26" toggle for filtering to current-semester offerings
• AI-generated answer with inline course code links to CAB
• Best Matches / Other Relevant Courses split — cited courses appear first, with additional results below
• Course cards showing sections, schedules, instructor, ratings, and truncated descriptions
• Concentration cards when a query matches a concentration program
• "What's being searched?" modal showing data coverage details