Similarity

Query

The results do not always contain all the query terms: a study can rank first even if only some of the words are present, as long as those terms are highly relevant.

Example: cryoglobulinemia vasculitis double-blind study

Preprocess

Each query goes through a preprocessing pipeline to normalize and clean the data before searching.

The main steps are:

1. Unicode normalization – remove accents and special characters.
2. Acronym expansion – replace acronyms (both generic like vs → versus and study-specific ones) by their definitions.
3. Roman numeral handling – convert expressions such as Grade IV or Phase II into numeric equivalents (Grade4, Phase2).
4. Lemmatization – reduce words to their base form (e.g., studies → study, blinded → blind).
5. Stopword – filter out common words (the, and, or, ...)
6. Digit and noise cleaning – remove standalone numbers, patterns like d14, w2, m6, and measurement expressions (10mg, 3ml, …).
7. Final filtering – lowercase everything, remove punctuation, and return a clean text.

This ensures that queries and documents are normalized, making matches more accurate and robust across variations in writing.

For our example :

You have to validate to display results

Display

The results are displayed as:

1. A score graph showing the top-ranked protocols for your query.
2. A list of matching studies with their titles and metadata.
3. Downloadable PDFs – download the protocol in PDF format.

For each study, you can open the document and view the highlighted occurrences of your search terms: