Database / Weaviate Vector database Interview questions
What is the difference between rescoring and raw compressed-vector search?
Raw compressed-vector search compares the query against every candidate using only the compressed (lossy) representation throughout, which is fast but accepts whatever recall degradation the compression technique introduces without any correction step. Rescoring adds a second pass: after an initial fast search over compressed vectors produces a candidate shortlist, Weaviate re-computes distances for just that shortlist using the original, full-precision vectors.
| Raw compressed search | Rescoring |
| Uses only compressed vectors throughout. | Uses compressed vectors for the initial pass, full-precision for the final ranking. |
| Fastest, but accepts full quantization recall loss. | Slightly slower (a small second pass), but recovers much of the lost recall. |
| No dependency on retaining uncompressed vectors. | Requires storing the original uncompressed vectors alongside the compressed ones. |
Because the second, rescoring pass only has to compute exact distances for a small shortlist (not the entire dataset), it's dramatically cheaper than a full uncompressed search would be, which is exactly why rescoring is such an effective technique: it gets most of compression's memory savings on the broad first pass while getting most of full-precision search's recall accuracy on the narrow, cheap second pass.
More Related questions...
