Skip to main content

Cross-Encoder Reranking

After semantic retrieval, ToolStorePy reranks the candidate documents with a cross-encoder.

Why reranking exists​

Embedding retrieval is fast and good at coarse recall, but it can produce plausible near-matches. The cross-encoder adds a slower, more precise relevance pass over the top-k candidate set.

Current default model​

cross-encoder/ms-marco-MiniLM-L-6-v2

Decision rule​

  • build [query, document] pairs
  • predict a score for each pair
  • keep the highest-scoring document as the match

Result format​

The selected document is parsed back into:

  • tool_id
  • tool_name
  • tool_description
  • tool_git_link
  • score