Special operations on indexes

There are a few additional operations that are not supported for all types of indexes.

Reconstructing vectors from an index

The methods reconstruct and reconstruct_n reconstruct one or several vector(s) from an index, given their ids.

Example usage: test_index_composite.py

Supported for IndexFlat, IndexIVFFlat (call make_direct_map first), IndexIVFPQ (same), IndexPreTransform (provided the underlying transform supports it)

Removing elements from an index

The method remove_ids removes a subset of vectors from an index. It takes an IDSelector object that is called for every element in the index to decide whether it should be removed. IDSelectorBatch will do this for a list of indices. The Python interface constructs this efficiently.

NB that since it does a pass over the whole database, this is efficient only when a significant number of vectors needs to be removed.

Example: test_index_composite.py

Supported by IndexFlat, IndexIVFFlat, IndexIVFPQ, IDMap.

Range search

The method range_search returns all vectors within a radius around the query point (as opposed to the k nearest ones). Since the result lists for each query are of different sizes, it must be handled specially:

• in C++ it returns the results in a pre-allocated RangeSearchResult[https://github.com/facebookresearch/faiss/blob/master/AuxIndexStructures.h#L35] structure

• in Python, the results are returned as a triplet of 1D arrays lims, D, I, where result for query i is in I[lims[i]:lims[i+1]], D[lims[i]:lims[i+1]]

Supported by (CPU only): IndexFlat, IndexIVFFlat, IndexScalarQuantizer, IndexIVFScalarQuantizer.

Splitting and merging indexes

The methods:

• merge_from copies another index to this and deallocates it on-the-fly. You can use ivflib::merge_into for IndexIVFs wrapped in a pre-transform.

• copy_subset_to copies a subset of this codes to another index. Example usage: to build indexes on a GPU and move them to CPU afterwards

The functions are implemented only for IndexIVF subclasses because they are mainly interesting for large indexes.