Limits and troubleshooting#

Compiled backend unavailable#

This submodule requires compiled Annoy support. An import/build problem in scikitplot.annoy is different from a problem in the independent scikitplot.cexternals._annoy owner. Keep those build domains separate when diagnosing failures.

The repository maintenance snapshot also records a build-structure issue in the private Cython Annoy path involving a missing Tempita generator. That is a build and release concern; user code should not work around it by copying headers or private generator files.

Common mistakes#

• adding a vector with the wrong dimensionality;

• querying before a useful index has been built;

• loading an index with a different metric/dimension expectation;

• expecting approximate neighbors to equal an exact search for every query;

• deleting or replacing a file while another process still memory-maps it; and

• using pickle with untrusted input.

When debugging retrieval quality, first verify vector generation and metric choice, then measure Annoy recall. Do not compensate for a broken embedding pipeline by increasing search_k indefinitely.