Brute force search without an index

In the case of IndexFlat, the added data is just copied to the index without further encoding or organization. Therefore, it may be useful to short-circuit the indexing altogether. This makes it possible to:

• avoid having two copies of the same data in memory

• do updated on the data between searches.

Faiss provides low-level functions to do the brute-force search in this context.

The functions take a matrix of database vectors and a matrix of query vectors and return the k-nearest neighbors and their distances.

Brute force search on CPU

On CPU, the relevant function is knn_L2sqr or knn_inner_product.

Example usage: TODO

Brute force search on GPU

On GPU, the relevant function is bruteForceKnn. An additional advantage is that it takes both CPU and GPU resident data as input or output.

An example usage in python, from pytorch GPU data is here: test_pytorch_faiss.py