← All work

DELIVERABLE 2 — WORKING RAG APPLICATION / AI Engineering / Retrieval Systems

Grounded RAG Policy Assistant

A retrieval-augmented question-answering assistant grounded in institutional policy documents with traceable citations and automated evaluation.

Digital vector grid and algorithmic network illustration representing semantic search
Deliverable 2 — Working RAG Application

This is Deliverable 2 for IS373: Individual RAG application. Below is the technical problem statement, retrieval architecture, evaluation benchmark suite, and system boundaries. An interactive live retrieval demonstration is available.

01

The question

Students and staff need direct, verifiable answers to campus policies without suffering hallucinations or manually hunting through dense PDF documents. Naive LLM prompts frequently invent rules or cite non-existent committees.

02

The process

Authored 5 synthetic policy documents (membership, event hosting, code of conduct, committee responsibilities, and directory). Designed a chunking pipeline splitting text into paragraph units with stable source IDs. Built vector similarity search with strict citation guarantees and an unanswerable query benchmark.

03

The outcome

Working Deliverable 2 application running at /rag. Features an automated 5-question evaluation bench testing ground-truth recall and an explicit 'insufficient evidence' fallback when questions fall outside the corpus (100% test pass rate).

04

The reflection

How do chunk boundary overlaps affect context preservation, and what server-side safeguards prevent unauthorized document ingestion?

Explore visual themes in the studio ↗