Information Retrieval System for Medical Document Analysis

Coursework for the Information Retrieval course at the University of Patras.
The project builds and evaluates several document-ranking models on the
Cystic Fibrosis medical literature collection (1,239 articles, 20 queries
with ground-truth relevance judgments), then explores document clustering with
both TF-IDF and semantic embeddings.

The code lives in Jupyter notebooks — there is no installable package or CLI.

What's inside

Notebook	What it does
`testing.ipynb`	Custom TF-IDF implementation (three TF weightings) + per-query evaluation (Precision@k, Recall@k, F1@k, MAP).
`vectorizer.ipynb`	scikit-learn `TfidfVectorizer` with a 270-combination grid search over the weighting/normalisation parameters.
`kmeans-vect.ipynb`	K-Means / agglomerative clustering on the TF-IDF vectors (elbow + silhouette analysis).
`embeddings.ipynb`	Semantic clustering with `sentence-transformers` (`all-MiniLM-L6-v2`, 384-dim).

Data files: Queries.txt (20 queries), Relevant.txt (relevance judgments),
and docs/ (the document collection).

Results

Retrieval (custom TF-IDF vs. grid-searched scikit-learn TF-IDF):

Model	F1@k	Precision	Recall	MAP
Raw TF-IDF	0.2817	0.0392	0.9009	0.2620
Log TF-IDF	0.2817	0.0392	0.9009	0.2616
Aug TF-IDF	0.2753	0.0392	0.9009	0.2583
Optimised TF-IDF	0.3006	0.0420	0.9200	0.2750

Best parameters: ngram_range=(1,1), sublinear_tf=True, min_df=1, max_df=0.7, norm='l2'.

Clustering:

Method	Best k	Silhouette
K-Means (TF-IDF)	17	0.18
Agglomerative (TF-IDF)	50	0.22
K-Means (embeddings)	12	0.31

Takeaways. On these short abstracts (~81 words avg, most terms appearing
≤2×) plain raw TF works as well as log/augmented weighting, and grid search
mainly helped by tuning max_df to drop over-common terms (F1@k 0.282 →
0.301). For clustering, sentence-transformer embeddings separated documents
noticeably better than TF-IDF vectors (silhouette 0.31 vs 0.18–0.22).

Running it

git clone https://github.com/thonos-cpu/TFIDF-Search-Engine.git
cd TFIDF-Search-Engine

python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate
pip install numpy scikit-learn sentence-transformers yellowbrick matplotlib jupyter

jupyter notebook                # then open any of the .ipynb files

Run testing.ipynb for the custom TF-IDF + evaluation, vectorizer.ipynb for
the grid search (a few minutes), and kmeans-vect.ipynb / embeddings.ipynb
for clustering. The embedding notebook downloads the all-MiniLM-L6-v2 model on
first run.

Course

Department of Computer Engineering & Informatics, University of Patras —
Winter 2025–2026. Supervisor: Prof. C. Makris. TAs: N. Kalogeropoulos,
A. Bompotas.

References

Shaw, Wood, Wood & Tibbo (1991) — Cystic Fibrosis test collection.
Salton & Buckley (1988) — Term-weighting approaches in automatic text retrieval.
Reimers & Gurevych (2019) — Sentence-BERT.

License

MIT — see LICENSE.

TFIDF Search Engine