Semantic comparison
Compare embedding vectors for two pieces of text or other embedded values.
AI Tools · Function
cosineSimilarity
Calculates the cosine similarity between two numeric vectors. Its signed result ranges from -1 to 1, with values near 1 indicating greater similarity.
How it works
cosineSimilarity compares two numeric vectors using the cosine of the angle between them. It returns a score from -1 to 1: values near 1 indicate strong similarity, while values near -1 indicate that the vectors differ.
Syntax
cosineSimilarity(vector1: number[], vector2: number[]): numberParameters
| Name | Type | Description |
|---|---|---|
vector1 | number[] | The first vector to compare |
vector2 | number[] | The second vector to compare |
Returns
A number between -1 and 1 representing the cosine similarity between the two vectors.
Key points
- The function measures similarity directly between two numeric vectors.
- Scores near 1 indicate that the compared vectors point in similar directions.
- Scores near -1 indicate that the vectors differ.
- The vector1 argument supplies the first vector.
- The vector2 argument supplies the second vector.
- Use the function after generating embeddings to compare represented values.
Common mistakes
- Do not interpret the result as a percentage; it is a signed score bounded by -1 and 1.
Example: Compare two vectors
import { cosineSimilarity } from 'ai';
const first = [1, 0, 0];
const second = [0.8, 0.2, 0];
console.log(cosineSimilarity(first, second));Compares two numeric vectors and prints their cosine similarity.
When to use it
Ranking related items
Order items by their similarity score to a reference embedding.
Filtering matches
Keep only dataset entries whose similarity score meets a chosen threshold.
Interview questions
QWhat does cosineSimilarity return?
It accepts two number arrays and returns their cosine similarity as a number. A result close to 1 means the vectors are very similar, whereas a result close to -1 means they are different.
Follow-up: How would you use it with generated embeddings? Embed the values first, then pass the resulting embedding arrays to cosineSimilarity.
QHow do you import and call the function?
The function is imported from the ai package. Its two inputs are vector1, the first vector to compare, and vector2, the second vector to compare.
Follow-up: What should the inputs represent? They should be numeric vectors whose values encode the items being compared, such as embeddings.
QWhen would you use this function?
It is useful after creating embeddings when comparing the semantic similarity of text or other represented values. The documentation demonstrates calculating a score between two embeddings.
Follow-up: What can be done with multiple scores? Similarity scores can support ranking or filtering related items in a dataset.
Related
embedManyembed