Update CAGRA docs (#4152)

Summary:
Update CAGRA train and add docs to reflect current capabilities.

Pull Request resolved: #4152

Reviewed By: mnorris11

Differential Revision: D68901974

Pulled By: asadoughi

fbshipit-source-id: 2667466286a48f7ed76017388f8bbd880025afdd

Author: tarang-jain

faiss/gpu/GpuIndexCagra.h

@@ -245,7 +245,11 @@ struct GpuIndexCagra : public GpuIndex {
             faiss::MetricType metric = faiss::METRIC_L2,
             GpuIndexCagraConfig config = GpuIndexCagraConfig());
 
-    /// Trains CAGRA based on the given vector data
+    /// Trains CAGRA based on the given vector data.
+    /// NB: The use of the train function here is to build the CAGRA graph on
+    /// the base dataset and is currently the only function to add the full set
+    /// of vectors (without IDs) to the index. There is no external quantizer to
+    /// be trained here.
     void train(idx_t n, const float* x) override;
 
     /// Initialize ourselves from the given CPU index; will overwrite

faiss/gpu/impl/CuvsCagra.cuh

@@ -118,8 +118,11 @@ class CuvsCagra {
     const int dim_;
 
     /// Controls the underlying cuVS index if it should store the dataset in
-    /// device memory
-    bool store_dataset_;
+    /// device memory. Default set to true for enabling search capabilities on
+    /// the index.
+    /// NB: This is also required to be set to true for deserializing
+    /// an IndexHNSWCagra object.
+    bool store_dataset_ = true;
 
     /// Metric type of the index
     faiss::MetricType metric_;