HuggingFace reranker

Reorders retrieved memories with a local HuggingFace-hub cross-encoder, selected via RERANKER_PROVIDER=huggingface. Reranking reorders retrieved memories by relevance after hybrid search.

Enable it

RERANKER_PROVIDER=huggingface
HF_RERANK_MODEL=BAAI/bge-reranker-base

Requires sentence-transformers: pip install sentence-transformers.

Configuration

Variable	Default	Description
RERANKER_PROVIDER	none	Set to huggingface.
HF_RERANK_MODEL	BAAI/bge-reranker-base	HuggingFace cross-encoder model to load.

Notes

• Runs locally via sentence-transformers, with the model loaded lazily on first use.
• Distinct from cross_encoder in that it defaults to a HF-hub reranker selectable via HF_RERANK_MODEL.
• The bge-reranker emits raw logits, which are sigmoid-mapped into [0, 1] to match the cross_encoder score scale.