Search Playground

Interactive demos of how search engines find and rank documents, from raw text to modern hybrid search. Every demo is live: edit the text, move the sliders, and watch the results change.

How does search work?

A search engine answers one question over and over: which documents best match this query? There are two main ways to answer it.

Keyword search looks for documents sharing words with the query. Not all words are equal: a document mentioning "quantum" several times is a better match for "quantum physics" than one full of "the". TF-IDF and BM25 turn this intuition into scores.

Semantic search compares meaning instead of words. Texts become vectors, and a search for "machine learning" can find a document about "neural networks" without a single word in common.

The demos follow the path of a real search engine. Each one stands on its own, but they make the most sense in order.

Preparing text
1

Text Analysis

Tokenization, stop words, synonyms, stemming and lemmatization: how raw text becomes searchable terms.

 2

Inverted Index

The data structure behind keyword search: from each term to the documents containing it.
Ranking by words
3

TF-IDF

Weigh each word by how often it appears in a document, and how rare it is in the collection.

 4

BM25

TF-IDF's successor, with term frequency saturation and document length normalization.
Ranking by meaning
5

Vector Search

Turn texts into vectors and rank documents by the angle between them and the query.

 6

Hybrid Search

Combine BM25 and vector rankings so each covers the other's blind spots.
Measuring quality
7

Evaluation

Precision, recall, reciprocal rank and nDCG: how to tell whether a ranking is good.