4 ms·
Try this trick that I learned from Cohere: - Fetch top 10*k (i.e. 100) results using the hamming distance - Rerank by taking dot product between query embedd
by amitness 2y ago
Try this trick that I learned from Cohere:
- Fetch top 10*k (i.e. 100) results using the hamming distance
- Rerank by taking dot product between query embedding (full precision) and binary doc embeddings
- Show top-10 results after re-ranking
- Quizzical4230 2y agoThis is pretty cool. The dot product would give the unnormalized cosine similarity from a smaller pool. Thank you so much!