Perplexity Embedding API Analysis: Unnormalized Vector Similarity and Dimension Reduction Effects

At a glance

An analysis of Perplexity's embedding API (a service that turns text into numerical vectors) found that the vectors come out unnormalized. This means you must specify how similarity is calculated to get accurate search results. The analysis also concluded that reducing the default 2,560 dimensions to 512 can still improve search speed and save storage space.

Key points

1. 1

   Reducing a 2,560-dimension vector to 512 dimensions greatly cuts storage needs and speeds up search.

2. 2

   The 8-bit integer method uses one byte per dimension, while the binary method uses only one bit per dimension.

3. 3

   Unnormalized vectors should be compared using cosine similarity (a value that measures how closely two vectors point in the same direction).

4. 4

   Binary vectors can be compared quickly using Hamming distance (a method that counts the number of differing bits).

Why it matters

People who build or run search features can learn which comparison method gives accurate results.