Double nlist, halve nprobes

* This achieves ~68% performance while cutting problem size in half

app.py

@@ -17,7 +17,7 @@ index = faiss.read_index('index.faiss')
 
 ps = faiss.ParameterSpace()
 ps.initialize(index)
-ps.set_index_parameters(index, 'nprobe=32,ht=512')
+ps.set_index_parameters(index, 'nprobe=16,ht=512')
 
 
 def _recover_abstract(inverted_index):
@@ -37,7 +37,7 @@ def search(query):
 
     query_embedding = model.encode(query)
     query_embedding = query_embedding.reshape(1, -1)
-    distances, faiss_ids = index.search(query_embedding, 10)
+    distances, faiss_ids = index.search(query_embedding, 20)
 
     distances = distances[0]
     faiss_ids = faiss_ids[0]
@@ -148,5 +148,5 @@ with gr.Blocks() as demo:
         .success(execute_request, inputs=[request_str_var], outputs=[response_var]) \
         .success(format_response, inputs=[neighbors_var, response_var], outputs=[results])
 
-demo.queue()
+demo.queue(2)
 demo.launch()

index.faiss

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:805bb6992b07e1c03395e82ef3dcbaf6624c7300450236a484b8d25ab30d2bff
-size 69337388
+oid sha256:bf114f0cc57e5674b171e113553e182aa705be7df530493862d21bb48b7ecf9b
+size 138019116

index.ivfdata

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0d30dada2642b662614cf6d539cb4de2c9f11785150fd9919fce574b8a3acfd3
+oid sha256:4d1bbbbb3094702c549d2a2576280a2587622241d0e7ec2765955a3461d89bae
 size 6868108080