CollegeAdvisor
An answer should come with a source.

A library with a gate on the way out.
An illustration built for this site in Blender. It explains the idea; it is not data from the project.
Admissions, financial aid, and transfer pathways all depend on specific information. At Athena AI Labs, I worked on an application that retrieves supporting material before constructing an answer, with checks for questions the available sources cannot support. The design goal was modest and strict: an answer should either show where it came from or decline.
Connecting the whole application
I developed a SwiftUI interface, a Python API, and retrieval-augmented generation workflows for college-advising questions. Student-profile integration supported personalized recommendations, while document ingestion organized the information the system could consult.
Five collections, kept apart
The knowledge base is organised as five collections: financial-aid policies, major admission requirements, Common Data Set figures, transfer articulation agreements, and worked answers with their citations. Keeping them separate lets a question about transfer credit search different material from a question about aid, and lets a result say which kind of source it came from.
Retrieval before generation
A question is first matched against those collections by meaning, and results from official .edu and .gov sources are weighted above other material. The published design documents also describe keyword retrieval; the path I describe here is the dense one. The language model’s job is to phrase the answer. The retrieved evidence supplies the facts.
Cite, calculate, or abstain
Some questions should not be answered from text at all. Aid and cost figures are computed by a deterministic calculator rather than generated. Questions that ask for legal judgement, predictions or personal preference are caught before retrieval. And when the best evidence is weak, the system abstains and explains why. On the project’s own stress set, most cases were resolved by declining to answer.
Making the work inspectable
I published CollegeAdvisor-RAG on Hugging Face with architecture notes and usage instructions. A model listing is also available on Kaggle. The portfolio focuses on the retrieval and citation design; benchmark claims in a model card should not be read as a guarantee for every admissions question.
Retrieval does the factual work.
- 01
Search by meaning, rank by authority
Each record is embedded and stored in one of five ChromaDB collections. A question retrieves candidates from every collection, and candidates that cite authoritative domains are boosted before the top results are kept. A minimum score keeps weak matches out of the answer.
- 02
Check the question and the source
Specialised handlers route questions about aid, transfer and admissions to the right material. A cite-or-abstain check keeps insufficient evidence visible rather than forcing every question into a confident answer. Student context enters through the application and API.
- 03
A smaller model, a narrower job
The published architecture uses a LoRA-adapted TinyLlama 1.1B model for response formatting. The retrieved evidence supplies factual content. Keeping those responsibilities separate makes it easier to inspect whether an answer is supported by its citations.
Architecture details follow the published model card and my own work on the project. They describe the system’s design, not a guarantee that every retrieved fact is current or every answer is correct.
Explore the model cardThe honest edges.
- The evaluations were written by the project itself. There is no independent benchmark or real-user accuracy figure.
- A citation shows where an answer came from. It does not prove the source is current.
- The Hugging Face release is a documented model artifact, not a hosted advisory service.
Follow the evidence.
- 01Question
Student context and a specific information need.
- 02Retrieve
Search across five source collections, weighting authoritative domains.
- 03Validate
Specialized handling and cite-or-abstain checks.
- 04Respond
An answer formatted with supporting citations.
Where this connects.
The diagram is a simplified presentation of the published architecture. The Hugging Face release is a documented model artifact, not a hosted advisory service.


