What is semantic search?

Semantic search finds content by meaning rather than exact keyword matching. It uses vector embeddings — numerical representations of text produced by a language model — to compare how similar two pieces of text are, even when they use different words.

What is a vector embedding?

A vector embedding is a list of numbers (a vector) that represents the meaning of a word, sentence, or document. Models like Sentence Transformers convert text into embeddings so that semantically similar texts produce vectors that are close together in high-dimensional space.

What is cosine similarity and why is it used in semantic search?

Cosine similarity measures the angle between two vectors, giving a value between -1 and 1. It is used in semantic search because it focuses on direction (meaning) rather than magnitude (length), so a short and a long document that mean the same thing score equally. A cosine similarity near 1 means very similar meaning.

What is the difference between semantic search and keyword search?

Keyword search matches documents that contain the exact words in the query. Semantic search matches documents by meaning, so it can find “how to fix flat tyre” when you search for “bike puncture repair”. Semantic search uses embeddings and vector similarity; keyword search uses term frequency and BM25 scoring.

Which Python library is best for semantic search?

The sentence-transformers library (by SBERT) is the most popular Python library for semantic search. It provides pre-trained models that convert sentences into high-quality embeddings. For storage and retrieval at scale, pair it with a vector database such as Faiss, Qdrant, Pinecone, or Weaviate.

Blog

Semantic Search & Vectors

Vector visualizer + intuition for embeddings, dot product, and cosine similarity.

Why do we need semantic search?

Traditional search (keyword matching) finds documents that contain the exact words the user typed. That fails when the user uses different words or phrasing. Semantic search finds content by meaning — it can match related concepts even when the words differ.

Vectors (embeddings) are the technical tool that let us compare meanings numerically.

How vectors represent meaning

An embedding is a vector of numbers created by a model to represent the meaning of text. Two pieces of text with similar meaning should have vectors that are close in the vector space.

Simple 2D illustration (for intuition):

cat → [1, 2]
dog → [2, 2] (close to cat → similar meaning)
car → [9, -1] (far away → different meaning)

cat [1,2]

dog [2,2]

car [9,-1]

Dot product & cosine similarity — visual intuition

The math behind semantic search with a wide horizontal graph showing the cat, dog, and car vectors side-by-side.

Dot product (a·b) measures raw overlap — both direction and length matter. Cosine similarity divides by lengths to focus only on direction (the angle θ between arrows).

Formula (short): a·b = Σ a_i b_i

Cosine: cosine(a, b) = (a·b) / (‖a‖ · ‖b‖) = cos(θ)

Using our simple 2D examples (intuition + toy numbers):

cat vs dog → arrows point in a very similar direction → cosine ≈ 0.95 (high similarity)
cat vs car → arrows point very differently → cosine ≈ 0.34 (low similarity)

Real-world / practical examples

1) Synonyms / paraphrases: "How to fix bike puncture" vs "How to repair a flat tyre" — their embeddings are very close directionally even if the words differ. Toy 4-dim vectors:

q1 = [0.20, 0.10, 0.90, 0.05]
q2 = [0.19, 0.11, 0.88, 0.06]
cosine(q1,q2) ≈ 0.99 → very similar

2) Length bias avoided: a long document may have a larger embedding norm. Dot product would favor the long doc even if meaning is the same. Cosine normalizes length so both short and long documents that mean the same score equally.

q = [0.5, 0.5, 0.1]
d_short = [0.5, 0.49, 0.11]
d_long = [5.0, 4.9, 1.1]  (same direction scaled)
// dot(q,d_short)=0.506, dot(q,d_long)=5.06  (dot prefers long)
// cosine(q,d_short) ≈ cosine(q,d_long) ≈ 0.999  (cosine treats them equal)

In practice: many vector DBs store normalized embeddings or compute cosine on the fly (Faiss, Milvus, Pinecone). That lets systems rank by semantic similarity reliably rather than by raw magnitude.

Imagine each sentence as an arrow. Semantic search finds arrows that point in the same direction — those are the meanings that match.

3D Vector Space (Interactive Demo)

Add words one by one and watch how vectors position themselves in 3D. Try similar words to see them cluster.

Enter one word (no spaces, max 50 chars):

0/50

Members: none

Drag to rotate · Scroll to zoom

Comments

Ask questions, share feedback, or reply to other readers. Sign up with Google to join the discussion.

Loading comments...