[ENHANCEMENT] Relevance limit settings

[dvejsada]

Consider possibility to add relevance cutoff as a setting via env. variable. If case of larger knowledge bases, file search returns quite a lot of document chunks even with quite little relevance score. If you use some of the larger models (e.g. GPT-4o or Clause 3.5 Sonnet), it can increase costs quite quickly.

Am I correct to assume that the lower the relevance score the better? Usually in our searches, the actually relevant documents has around 0.2-0.3 relevance score, the least relevant 0.7 - 0.8. In total it return about 25 pages of text, which is quite lot.

[dvejsada]

After some additional thought, it would be beneficial to have a default relevance limit with optionally being able to overwrite this default for specific endpoint/agent in LibreChat. In some use cases you want only very relevant results, in others broader scope of results e.g. for additional information.