adviSU — Retrieval-Augmented Academic Advising System

Overview

A bilingual (Turkish/English) academic advising assistant that answers degree, course-selection, and scheduling questions from official university curriculum data. Built as a Sabancı University CS455 course project.

Problem

Academic advising questions — can I graduate on time, which courses am I eligible for, does this timetable conflict — depend on exact curriculum rules and personal course history. A language model asked these questions directly will hallucinate credit counts, prerequisites, and schedule conflicts, which is unacceptable when students make degree decisions from the answers.

Solution

A deterministic-first pipeline: graduation arithmetic, prerequisite checks, and schedule conflicts are computed in code, and the language model is restricted to explaining verified results. A hybrid retrieval layer grounds every answer in official curriculum documents, and a confidence policy makes the system abstain rather than guess when evidence or profile data is insufficient.

Highlights

• •Deterministic academic engines for degree auditing, prerequisite- and stage-aware course recommendation, conflict-free weekly timetable construction, GPA projection, and major selection
• •Hybrid retrieval layer combining field-weighted BM25F, dense multilingual embeddings, reciprocal rank fusion, and cross-encoder reranking — scoped by program and curriculum term before ranking
• •Provider-neutral LLM adapter over Groq, Mistral, and OpenRouter with exact model pinning, bounded retries, circuit breaking, and sanitized error handling
• •Layered safety controls: content-safety classification, prompt-injection and secret-extraction guardrails, retrieval quarantine, and output validation
• •Confidence policy that abstains rather than guessing when evidence or profile data is insufficient
• •HMAC-keyed, fail-closed resource controls for per-user quota, token, cost, and concurrency limits across multi-worker deployments
• •React front end with streaming chat, profile and course-history management, degree audit, schedule workspace, spreadsheet export, and a single typed localization catalogue
• •Reproducible evaluation framework with preregistered method selection, frozen benchmark splits, leakage-controlled data partitioning, and checksummed artifacts

Tech Stack

Links