Uploaded image for project: 'MariaDB Server'
  1. MariaDB Server
  2. MDEV-36317

vector search with Cosine Distance, the recall rate of the returned results is very low

Details

    • Bug
    • Status: Closed (View Workflow)
    • Critical
    • Resolution: Won't Fix
    • None
    • N/A
    • Vector search
    • None

    Description

      vector search with Cosine Distance, the recall rate of the returned results is very low.
      abs2 in FVector is memory value, when we load data from disk, we need init abs2 = 1.0f
      The fix code is as follows

      diff --git a/sql/vector_mhnsw.cc b/sql/vector_mhnsw.cc
      		 
      index d8a63a7558c..91256a31910 100644
      		 
      --- a/sql/vector_mhnsw.cc
      		 
      +++ b/sql/vector_mhnsw.cc
      		 
      @@ -820,7 +820,7 @@ int FVectorNode::load_from_record(TABLE *graph)
      		 
         FVector *vec_ptr= FVector::align_ptr(tref() + tref_len());
      		 
         memcpy(vec_ptr->data(), v->ptr(), v->length());
      		 
         vec_ptr->postprocess(ctx->vec_len);
      		 
      -
      		 
      +  if (ctx->metric == COSINE) vec_ptr->abs2 = 1.0f;
      		 
         longlong layer= graph->field[FIELD_LAYER]->val_int();
      		 
         if (layer > 100) // 10e30 nodes at M=2, more at larger M's
      		 
           return my_errno= HA_ERR_CRASHED;

      Attachments

        Issue Links

          blocks

          New Feature - A new feature of the product, which has yet to be developed. MDEV-33408 HNSW for k-ANN vector searches

          • Critical - Crashes, loss of data, severe memory leak.
          • Closed
          relates to

          Bug - A problem which impairs or prevents the functions of the product. MDEV-36313 VEC_DISTANCE_COSINE return error value

          • Minor - Minor loss of function, or other problem where easy workaround is present.
          • Closed

          Activity

            Ascending order - Click to sort in descending order
            serg Sergei Golubchik added a comment -

            Do you have a test case? Normally vec_ptr->abs2 should be 1.0f here already. What values of vec_ptr->abs2 do you see?

            serg Sergei Golubchik added a comment - Do you have a test case? Normally vec_ptr->abs2 should be 1.0f here already. What values of vec_ptr->abs2 do you see?
            myx myx added a comment -

            I run ann bench mark with dataset nytimes-256-angular

            myx myx added a comment - I run ann bench mark with dataset nytimes-256-angular
            serg Sergei Golubchik added a comment -

            Yes, I've run ann-benchmarks with nytimes-256-angular. Is this patch your way of fixing zero-length vectors? Does it help to improve recall?

            serg Sergei Golubchik added a comment - Yes, I've run ann-benchmarks with nytimes-256-angular. Is this patch your way of fixing zero-length vectors? Does it help to improve recall?
            myx myx added a comment -

            yes, it significantly increased the recall rate

            myx myx added a comment - yes, it significantly increased the recall rate
            myx myx added a comment -

            Because all elements in the vector are 0, abs2 is also 0，and when we load the vector from disk，postprocess funciton set abs2 = 0

            myx myx added a comment - Because all elements in the vector are 0, abs2 is also 0，and when we load the vector from disk，postprocess funciton set abs2 = 0
            serg Sergei Golubchik added a comment -

            I've tested your suggested change.

            It has no effect on recall. Unexpectedly for me it affected the qps (the search speed). For low recall (<80%) it made the search slightly faster, for higher recall — somewhat slower. Because the effect was not very big (few %% faster, not few times) and recalls below 80% aren't practically useful anyway, I'm not going to apply this change for now.

            serg Sergei Golubchik added a comment - I've tested your suggested change. It has no effect on recall. Unexpectedly for me it affected the qps (the search speed). For low recall (<80%) it made the search slightly faster, for higher recall — somewhat slower. Because the effect was not very big (few %% faster, not few times) and recalls below 80% aren't practically useful anyway, I'm not going to apply this change for now.

            People

              serg Sergei Golubchik
              myx myx
              Votes:
              0 Vote for this issue
              Watchers:
              3 Start watching this issue

              Dates

                Created:
                Updated:
                Resolved:

                Git Integration

                  Error rendering 'com.xiplink.jira.git.jira_git_plugin:git-issue-webpanel'. Please contact your Jira administrators.