nearVector search on 3d vector space returns unexpected results

Support
1

Hi there. I am completely new to Weaviate and experimenting with loading and querying some simple objects. Each object has properties x, y, and z, each of which is a number, and a 3d vector consisting of these same numbers.

Here is my schema:

{
      'class': "XYZ",
      'description': 'A 3d vector class',
      'vectorizeClassName': false,
      'vectorizer': 'none',
      'properties': [
        {
          'name': 'x',
          'dataType': [ 'number' ],
          'description': 'x',
          'vectorizePropertyName': false
        },
        {
          'name': 'y',
          'dataType': [ 'number' ],
          'description': 'y',
          'vectorizePropertyName': false
        },
        {
          'name': 'z',
          'dataType': [ 'number' ],
          'description': 'z',
          'vectorizePropertyName': false
        },
      ]
    }

I load 1000 randomly-generated objects, and they seem to load fine:

curl -s "http://localhost:8080/v1/objects?class=XYZ&limit=1&include=vector" | jq .
{
  "deprecations": [],
  "objects": [
    {
      "class": "XYZ",
      "creationTimeUnix": 1689563219367,
      "id": "00d3baec-56ec-43eb-b860-4958afc8da07",
      "lastUpdateTimeUnix": 1689563219367,
      "properties": {
        "x": 178,
        "y": 840,
        "z": 330
      },
      "vector": [
        178,
        840,
        330
      ],
      "vectorWeights": null
    }
  ],
  "totalResults": 1
}

However, I find that regardless of the vector property that I pass in my nearVector query, I always get the same top-n results back:

echo '{
    "query": "{
      Get{
        XYZ(
          nearVector: {
            vector: [100, 100, 100]
          }
          limit: 1
        ){
          x
          y
          z
        }
      }
    }"
  }' | curl -s -X POST -H 'Content-Type: application/json' -d @- http://localhost:8080/v1/graphql | jq .
{
  "data": {
    "Get": {
      "XYZ": [
        {
          "x": 454,
          "y": 552,
          "z": 508
        }
      ]
    }
  }
}

and

echo '{
    "query": "{
      Get{
        XYZ(
          nearVector: {
            vector: [800, 800, 800]
          }
          limit: 1
        ){
          x
          y
          z
        }
      }
    }"
  }' | curl -s -X POST -H 'Content-Type: application/json' -d @- http://localhost:8080/v1/graphql | jq .
{
  "data": {
    "Get": {
      "XYZ": [
        {
          "x": 454,
          "y": 552,
          "z": 508
        }
      ]
    }
  }
}

Can someone help point me to where I’m going wrong?

Thanks!
–Scott

2

Hi, the HNSW is not good with random data. If you want to test it you need something more realistic such as the SIFT dataset: http://corpus-texmex.irisa.fr/

3

Thanks. Can you recommend a good resource on what the “shape” of the vectors/dataset needs to be to get effective results? All real-world data has some randomness in it, so I’m wondering if there are degrees of randomness that are acceptable?

4

Can’t answer that in general, but I ran into a similar problem when writing a test recently and what I did here worked