Question 1

How does this tool calculate similarity between text documents?

Accepted Answer

The tool provides two complementary modes: (1) an instant on-device sublinear TF-IDF lexical cosine engine, and (2) an in-browser neural sentence transformer engine (MiniLM-L6-v2) that maps text into 384-dimensional dense semantic vectors using WebAssembly SIMD.

Question 2

What is the difference between lexical TF-IDF and dense neural embeddings?

Accepted Answer

Lexical TF-IDF relies on exact word overlap and token frequencies. Dense neural embeddings project sentences into a continuous mathematical embedding space where semantic synonyms (e.g., "cat sat on the mat" and "feline rested upon the rug") cluster closely together even with zero common words.

Question 3

Why do two sentences with similar meanings sometimes score 0.000 in lexical mode?

Accepted Answer

In lexical mode, similarity requires shared vocabulary tokens. If two sentences express identical ideas using distinct synonyms, their lexical dot product is 0. Switching to Neural Mode uses dense embeddings to recognize the conceptual equivalence.

Question 4

How is cosine similarity mathematically computed?

Accepted Answer

Cosine similarity equals the dot product of two vectors divided by the product of their Euclidean lengths (L2 norms): Similarity = (A · B) / (||A|| ||B||). Scores range from 0.000 (completely orthogonal) to 1.000 (identical direction).

Question 5

Are my search queries or private text documents sent to an external server?

Accepted Answer

No. Both the lexical math and the neural ONNX model execution run strictly within your local browser runtime. No text or embeddings ever leave your device.

Semantic Vector Similarity Studio

Client-side neural embeddings and vector similarity

Frequently asked

Feedback, bugs & custom requests

Related tools