A re-ranker model takes the retrieved documents from a retriever and the user query as an input, and provides a score to each document based on their relevance to the query. This can then be used to "re-rank" the documents.

The score provided by the re-ranker is theoretically better than the score we already receive from the retriever, which gives the score fetched from the similarity score. But the re-ranker score would be more relevant to the user query. Using faster re-ranker models can also be quick in providing results.

If the re-ranker performs well, we can use a score threshold (as used previously) and can potentially omit the need for the relevance check.

This can be investigated directly with the current retrieval setup, or using opensearch