UC Davis Course Advisor

A RAG-powered bilingual course advisor for UC Davis students, featuring dual-path retrieval, prerequisite DAGs, and Claude-generated recommendations.

Live demo: ucd-course-advisor.fly.dev [cold start: refresh after 2min if Error:502]

A RAG-powered course advisor that lets UC Davis students ask questions in English or Chinese and get personalized, structured course recommendations — aware of prerequisites, student level, and completed coursework.

Architecture

scrape-catalog.py  →  courses_raw.json
                              ↓
                       build_index.py          build_dag.py
                    (bge-m3 embeddings       (prerequisite DAG
                     → chroma_db/ +           → course_dag.pkl)
                       bm25_index.pkl)
                              ↓
                          RAGPipeline
                    vector + BM25 → RRF fusion
                    → reranker → level boost
                    → DAG tier annotation
                    → Claude API → answer

Retrieval: Dense vector search (bge-m3) and BM25 keyword search run in parallel, fused via weighted RRF (0.7/0.3), then re-scored by bge-reranker-v2-m3 with a level boost to surface appropriately-leveled courses.

DAG tier annotation: Every retrieved course is tagged based on the student’s completed coursework:

  • ✅ Available Now — all prerequisites met
  • ➡️ Coming Soon — 1–2 prerequisites missing
  • 📅 Long-term Plan — multi-course path needed

Generation: Retrieved context is passed to Claude, which produces a structured, readable answer grounded in actual catalog data.

Screenshots

Onboarding: students enter their major, completed courses, and level
Structured response for a Statistics major asking about AI courses

Key Design Decisions

BM25 + vector fusion — Pure vector search hallucinated on rare course names. Adding BM25 with keyword boosting significantly reduced irrelevant retrievals, especially for Chinese queries.

Prerequisite DAG — Built with OR/AND logic to handle complex prerequisite chains, visualizable per department.

No LLM judge in retrieval — Every retrieval scoring step uses deterministic signals. Claude is only called once at generation, keeping costs and latency low.

Bilingual support — Handles English and Chinese queries natively, with keyword expansion for mixed-language inputs.

Stack: Python · FastAPI · ChromaDB · bge-m3 · bge-reranker-v2-m3 · BM25 · Claude API · Fly.io