Leverage distance matrix API for Visualize tab

[coszio]

For the Visualize section, currently we do a random sampling of points, and pull them along with their vectors. Afterwards we process them all in the browser, depending on their dimensionality reduction algorithm.

Since Qdrant v1.12.0, we've had a dedicated distance matrix API, which should improve the algorithms' time of execution.

If the algorithms accept a precalculated matrix, it would be nice to make use of this API before passing the work to the browser.

[nleroy917]

Oh awesome I had no idea this existed. The EmbeddingAtlas component has support for neighbors to potentially add nice visualizations. I think this could be useful there.

As for the UMAP computation, I agree that would be nice -- it just depends if their algorithms API supports that, I think

[nleroy917]

Trying to grab the matrix pairs from the API using this call in JS:

/**
 * Fetch k-nearest neighbors for all points using Qdrant's batch search API
 * @param {QdrantClient} qdrantClient - Instance of QdrantClient
 * @param {string} collectionName - Name of the Qdrant collection
 * @param {Array} pointIds - Array of point IDs to fetch neighbors for
 */
export async function fetchNearestNeighborsFromQdrant(qdrantClient, collectionName, pointIds) {
  return qdrantClient.searchMatrixPairs(collectionName, {
    filter: {
      must: [
        {
          key: 'id',
          match: {
            any: pointIds,
          },
        },
      ],
    },
  });
}

Returns nothing though. Wondering if I am using the filter wrong? The visualization chart first grabs points via scroll like so:

/* eslint-disable camelcase */
export function requestData(qdrantClient, collectionName, { limit, filter = null, using = null, color_by = null }) {
  // Based on the input parameters, we need to decide what kind of request we need to send
  // By default we should do scroll request
  // But if we have color_by field which uses query, it should be used instead

  if (color_by?.query) {
    const query = {
      query: color_by.query,
      limit: limit,
      filter: filter,
      with_vector: using ? [using] : true,
      with_payload: true,
      using: using ?? null,
    };

    return qdrantClient.query(collectionName, query);
  }

  // It it's not a query, we should do a scroll request

  const scrollQuery = {
    limit: limit,
    filter: filter,
    with_vector: using ? [using] : true,
    with_payload: true,
  };

  return qdrantClient.scroll(collectionName, scrollQuery);
}

[nleroy917]

Ok more intersetingly, even a curl with no filters doesn't return anything... must be using this API wrong!

curl -X POST http://localhost:6333/collections/products/points/search/matrix/pairs \
     -H "api-key: <apiKey>" \
     -H "Content-Type: application/json" \
     -d '{}' | jq
  % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current
                                 Dload  Upload   Total   Spent    Left  Speed
100    58  100    56  100     2   2933    104 --:--:-- --:--:-- --:--:--  3052
{
  "result": {
    "pairs": []
  },
  "status": "ok",
  "time": 0.002679041
}

[nleroy917]

turns out i was missing "using": "dense" -- skill issue!

[nleroy917]

Ok, my next issue is that I am trying to get the distances for the points pulled in the sample. On the UI we can specify limit. We scroll for this many points and use them to visualize the UMAP. We can get the actual point ids:

const pointIds = result.points.map((p) => p.id);

I pass this to my query to fetch the distance matrix and use a filter to pull down the points:

curl -X POST http://localhost:6333/collections/products/points/search/matrix/pairs \
     -H "Content-Type: application/json" \
     -d '{
            "using": "dense",
            "sample": 1000,
            "filter": {
                "must": [{
                    "key": "id",
                    "match": {
                        "any": [1,2,3,4,5]
                    }
                }]
            }
        }' | jq

But this actually just results in no points being returned... I'm curious if I am doing something wrong. Is it possible to filter the matrix pairs like this?

[coszio]

Thanks for looking into this @nleroy917!

Regarding your latest question, the match_any filter is intended to look inside the payload. Since we want to filter by point ID, then the relevant condition is has_id.