How to perform batch query in Vector Search

Munna_Ram · 2024-07-18T11:19:46.496Z

I want to send a request for multiple (let n=5), query vectors to find top 100 matching documents.
I am sending requests for all of them one by one.

Is there any method by using that I can get the result in a single request by searching with ANN in parallel for each query.

Shane · 2024-07-18T14:00:08.098Z

Would it be possible to run $facet with multiple $vectorSearch stages?

Something like:

pipeline = [{'$facet':
    {
      'query1': [{'$vectorSearch': {...}],
      'query2': [{'$vectorSearch': {...}],
      ...
    }
}]
docs = list(coll.aggregate(pipeline))

Henry_Weller · 2024-07-18T14:29:09.732Z

Shane:

pipeline = [{'$facet':
    {
      'query1': [{'$vectorSearch': {...}],
      'query2': [{'$vectorSearch': {...}],
      ...
    }
}]
docs = list(coll.aggregate(pipeline))

@Shane unfortunately $vectorSearch cannot be run as a pipeline within $facet so this wouldn’t work. I just tested this and you get:

pymongo.errors.OperationFailure: $vectorSearch is not allowed to be used within a $facet stage, full error: {'ok': 0.0, 'errmsg': '$vectorSearch is not allowed to be used within a $facet stage', 'code': 40600, 'codeName': 'Location40600', '$clusterTime': {'clusterTime': Timestamp(1721312815, 1), 'signature': {'hash': b'zl\xbd\x87\xb9\xb1\x00\xcf\xf8QS\x99Z;\xe8\xcb\xbe\x91\xd4\xe1', 'keyId': 7359286720640057346}}, 'operationTime': Timestamp(1721312815, 1)}

@Munna_Ram can you share more details about the different types of query vectors and what they represent and what indexed fields you’d like to compare them to?

Munna_Ram · 2024-07-20T11:47:31.017Z

@Henry_Weller
In my database I have documents with “DocEmbedding” as vector field with 768 dimensions, and some other fields like id, text, etc.

I want to query top 10 similar documents for the input query.
I have 10 different queries and I want to get results for them.
As they are independent then I want to perform this task in parallel for all 10 queries instead of sending requests in a loop to improve the performance of my application. I can do it by sending requests with different threads to get results in parallel but creating threads is a costly task.

So is it possible to execute $VectorSearch pipeline for each query independently and then return results for all queries in a list?

Munna_Ram · 2024-07-20T11:47:54.451Z

image954×370 21 KB

I think It can not be done with facet.
Run Vector Search Queries - MongoDB Atlas