Applied AI Case Study · 2026

Evidence Research Copilot

A small, inspectable research assistant built to make retrieval quality and evidence tracing visible instead of treating RAG as a black box.

Demo Recall@K1.004-case synthetic regression set, K=3

Demo MRR1.00Reproducible locally

Citation validity1.00Source IDs verified against retrieved chunks

Problem

RAG demos often show an answer, but hide why the system trusted it.

I wanted the portfolio project to expose the retrieval path directly: lexical rank, semantic-style rank, fused rank, selected evidence and citations.

The offline core avoids an API dependency so a reviewer can reproduce the retrieval and evaluation behavior without credentials.

System

documents→chunking→BM25 + TF-IDF→RRF→grounded synthesis→citations

Hybrid retrieval

Exact terminology is handled by BM25 while TF-IDF cosine contributes a second independent ranking signal.

Rank fusion

Reciprocal-rank fusion combines ranked lists without pretending the raw scores are calibrated to the same scale.

Evaluation

A committed evaluation set measures retrieval recall, MRR and citation validity on every regression run.

Engineering choice

The retrieval and fusion logic is implemented directly so I can explain each stage. FastAPI is only the interface layer.

The project takes inspiration from evaluation-first RAG tooling such as Ragas, but does not copy Ragas source code or assets.