You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
get_relevant_chunks gets sets of relevant passages via various retrieval methods (semantic/dense, sparse-embedding based, lexical/keyword, fuzzy); most of these produce scores (except fuzzy, I think), but they all are at least ordered by match quality.
If the config has a cross-encoder model set, then the union of these passages is then re-ranked and then finally the top $k$ passages are picked.
However if there's no cross-encoder, then the top $k$ are picked from this union of passages, which has no particular order, and we can get a bad set of passages. In this case we should make use of the scores (and implicit rank) that are produced by the retrievals, and blend them using a suitable fusion-ranking strategy, where the first step is to normalize the scores/ranks within each retrieval-set, to put them into the same "score space", so that picking top k is meaningful. The normalization could be:
reciprocal rank based -- score each passage by 1/(rank + c),
min-max-normalized
The text was updated successfully, but these errors were encountered:
get_relevant_chunks
gets sets of relevant passages via various retrieval methods (semantic/dense, sparse-embedding based, lexical/keyword, fuzzy); most of these produce scores (except fuzzy, I think), but they all are at least ordered by match quality.If the config has a cross-encoder model set, then the union of these passages is then re-ranked and then finally the top$k$ passages are picked.
However if there's no cross-encoder, then the top$k$ are picked from this union of passages, which has no particular order, and we can get a bad set of passages. In this case we should make use of the scores (and implicit rank) that are produced by the retrievals, and blend them using a suitable fusion-ranking strategy, where the first step is to normalize the scores/ranks within each retrieval-set, to put them into the same "score space", so that picking top k is meaningful. The normalization could be:
The text was updated successfully, but these errors were encountered: