Module resolver 

Context-aware resolver for face retrieval.

Sits between retrieve_candidates and final assignment: re-scores the ranked clusters using contextual signals and returns a stronger recommendation, or flags that the decision should still go to the user.

Signals fused:

• Co-occurrence: if the photo already has other assigned clusters, candidates with higher co-occurrence to those clusters are boosted. “Same event, same cast” principle.

• Temporal neighbors: if nearby-in-time photos have the candidate cluster already assigned, boost. “Same session → same people”.

The embedding score is always the dominant signal; context only breaks ties and nudges ambiguous cases.

Structs

ResolverContext

Context provided at resolve time. All queryable from FaceRepo.

ResolverWeights

Numeric weights controlling the relative influence of each signal. Default values roughly chosen so context nudges but doesn’t dominate a strong embedding disagreement.

Functions

rerank

Apply contextual signals to a ranked list of retrieval hits and return the re-sorted list.