How to use
- Paste vector A and vector B — as a JSON array, comma or space separated numbers, or a whole embedding API response (the first numeric array is extracted).
- Cosine similarity is shown large at the top, with distance metrics below.
- Use “Normalize A” to convert a vector to unit length, as many vector databases expect.
FAQ
Which metric should my vector database use?
Use what the embedding model was trained for — almost always cosine. If the vectors are already normalized (length 1), dot product gives the same ranking and is faster, which is why many databases default to it.
What is a “good” cosine similarity?
It depends heavily on the model. Some models put unrelated text around 0.1, others around 0.7. Compare scores from the same model against known similar and dissimilar pairs rather than using a fixed threshold.
Can I compare vectors from different models?
No. Vectors from different models (or different dimensions) live in different spaces, so their similarity is meaningless — the tool will refuse vectors of different lengths.