In the Faiss engine, when no filtering clause is applied, the search initially considers all available documents in a segment, and deleted documents are filtered out afterward.
This approach affects both recall and latency when there are many deleted documents. For ANN search, recall decreases because the results may include deleted documents among the top k. For exact search, latency increases due to the additional computation of distances for deleted documents.
This applies only with Faiss engine. With Lucene engine the live document information is already being used during search.
In the Faiss engine, when no filtering clause is applied, the search initially considers all available documents in a segment, and deleted documents are filtered out afterward.
This approach affects both recall and latency when there are many deleted documents. For ANN search, recall decreases because the results may include deleted documents among the top k. For exact search, latency increases due to the additional computation of distances for deleted documents.
This applies only with Faiss engine. With Lucene engine the live document information is already being used during search.