Configurable Reference Source Limit for Search Optimization

[goodmaney]

When using the native search feature in commercial models, I observed that pulling ~50 references produces the best output quality. However, this may be excessive for local models with limited context windows.

Proposed Solution:

Add a user-configurable setting for reference source limits to optimize for:

Commercial APIs (e.g. GPT-4, Claude): Higher limits (~50 sources) for accuracy
Local models: Lower limits to prevent context window overflows