MatchingEngineIndexEndpoint.read_index_datapoints drops numeric_restricts for private endpoints #7272

Environment details

• OS type and version: Linux
• Python version: 3.12
• google-cloud-aiplatform version: 1.130.0; the same code is on main at the time of writing

Summary

For an index deployed to a private (PSC / VPC) endpoint, MatchingEngineIndexEndpoint.read_index_datapoints() returns datapoints without numeric_restricts. Token restricts and crowding_tag are returned.

Steps to reproduce

1. Deploy an index to a private endpoint.
2. Upsert a datapoint that has a numeric restrict, e.g. NumericRestriction(namespace="expires_at", value_int=...).
3. Read it back with read_index_datapoints().

Code example

endpoint = aiplatform.MatchingEngineIndexEndpoint(index_endpoint_name=ENDPOINT)
endpoint.private_service_connect_ip_address = PSC_IP  # private endpoint

dp = endpoint.read_index_datapoints(deployed_index_id=DEPLOYED_INDEX_ID, ids=[DATAPOINT_ID])[0]
print(dp.restricts)          # token restricts are present
print(dp.numeric_restricts)  # empty, expected the numeric restrict

emb = endpoint._batch_get_embeddings(deployed_index_id=DEPLOYED_INDEX_ID, ids=[DATAPOINT_ID])[0]
print(emb.numeric_restricts)  # present: the server returns it

Cause

In the private-endpoint branch of read_index_datapoints (L1942-L1972), each Embedding from _batch_get_embeddings is converted to IndexDatapoint with feature_vector, restricts and crowding_tag only. Embedding.numeric_restricts (match_service.proto L138) is not copied.

MatchNeighbor.from_embedding already converts numeric_restricts for find_neighbors (L310), so this looks like an omission in the read path. The fix would map value_int / value_float / value_double to IndexDatapoint.NumericRestriction in the same way.

I'm happy to send a PR for this.