Embedding Similarity Explorer: Cosine, TF-IDF and Retrieval

Datalad tool

Embedding Similarity Explorer

Retrieval comes down to turning text into numbers and measuring the angle between them. This does that in front of you: real vectors, real cosine similarity, and the per word arithmetic behind every score. It also shows the exact point where matching on words alone falls over, which is the reason dense embeddings exist at all.

Your documents and your query

Documents, one per line

Query

The vectors are TF-IDF, built with the same formula scikit-learn uses, so the numbers here are the numbers you would get in a notebook. The geometry is identical to the geometry a dense embedding model uses. What differs is where the numbers come from, and that difference is the whole point of the example loaded below.

How the documents rank

Three ways of measuring the same vectors. They do not agree, and knowing why is most of what there is to know about retrieval scoring.

Where the top score came from

What this tells you

Tools

Andrei

August 17, 2026

1 min read

View Comments (5)

Recommended for You

Concurrent Test Collision Checker

Running several experiments at once is usually fine, for reasons people get wrong. See which of yours actually collide and how many users land in both.

August 17, 2026

1 min read

Bayesian vs Frequentist Comparator

The same result read both ways. A p value and the chance the variant is actually better answer different questions, and this shows both of them side by side.

August 17, 2026

1 min read

Python Data Structures Cheatsheet

Model Validation Cheatsheet

Matplotlib Cheatsheet

GitHub Workflow Cheatsheet