Course Compass
Course Compass is a college course advisor that runs on a student's own computer. It searches CHED and DepEd documents first, so the local language model only answers from what was found.
My part
I wrote documentation sections 4, 6, and 8, fixed parts of the other sections, and made the diagrams. The rest of this page describes the system those sections document.
The question the app answers
A student gives their strand, grades, interests, and career goal, and wants to know which college courses fit. A general chatbot can make things up, so the app first retrieves real passages from official CHED and DepEd documents and then asks the model to answer using only those.
How a request flows
- Build the query. The profile is flattened into one sentence used only for searching.
- Retrieve. The query is embedded and matched against the stored documents, returning the closest chunks.
- Augment. The chunks and the profile go into a fixed prompt that tells the model to use only the retrieved context.
- Generate and parse. The model replies with a list, and every item is checked against a schema.
The layers
| Layer | What it does |
|---|---|
| Presentation | A plain HTML and JavaScript form that shows the ranked courses |
| API | A FastAPI app with health, recommend, and PDF upload endpoints |
| Orchestration | Runs the stages in order and is the only part that talks to both the vector store and the model |
| Embedding | A local embedding model, downloaded once |
| Vector store | A ChromaDB collection saved on disk |
| Local model | Ollama running on the same machine, with no API key |
Why it runs locally
Every cloud dependency was replaced with a local one. Settings such as the model live in a settings file, so a different machine can swap them without code changes.