Update app.py

app.py

@@ -16,7 +16,7 @@ from statistical_chunker import StatisticalChunker
 from optimum_encoder import OptimumEncoder
 from scipy.sparse import csr_matrix, save_npz, load_npz, vstack
 from langchain_community.document_loaders import WikipediaLoader, WebBaseLoader
-from fastembed import SparseEmbedding
+from fastembed import SparseEmbedding, SparseTextEmbedding
 from unstructured.partition.auto import partition
 import onnxruntime as ort
 from pymilvus import (
@@ -44,7 +44,7 @@ def query_hybrid_search(col: Collection, query: str, dense_model: OptimumEncoder
     data = query_sparse.values
     sparse_matrix = vstack(csr_matrix(data, query_sparse.indices, np.array([0, len(data)]), shape=(1, max_index + 1))).tocsr()
 
-    sparse_req = AnnSearchRequest(query_sparse_embeddings, "sparse_vector", {"metric_type": "IP"}, limit=10)
+    sparse_req = AnnSearchRequest(sparse_matrix, "sparse_vector", {"metric_type": "IP"}, limit=10)
     dense_req = AnnSearchRequest(query_dense_embeddings, "dense_vector", {"metric_type": "COSINE"}, limit=10)
 
     res = col.hybrid_search([sparse_req, dense_req], rerank=RRFRanker(), limit=3, output_fields=["text"])
@@ -122,7 +122,7 @@ def load_models_and_documents():
             cache_dir=os.getenv('HF_HOME')
         )
         
-        sparse_model = Bm42(
+        sparse_model = SparseTextEmbedding(
             'Qdrant/all_miniLM_L6_v2_with_attentions',
             cache_dir=os.getenv('HF_HOME'),
             providers=['CUDAExecutionProvider']
@@ -143,10 +143,15 @@ def load_models_and_documents():
     schema = CollectionSchema(fields, "")
     col = Collection(collection_name, schema)
 
-    sparse_index = {"index_type": "SPARSE_WAND", "metric_type": "IP"}
-    dense_index = {"index_type": "FLAT", "metric_type": "COSINE"}
+    sparse_index = {"index_type": "SPARSE_INVERTED_INDEX", "metric_type": "IP"}
+    dense_index = {"index_type": "HNSW", "metric_type": "COSINE"}
     col.create_index("sparse_vector", sparse_index)
     col.create_index("dense_vector", dense_index)
+
+    col.alter_index(
+        index_name=("dense_vector"),
+        extra_params={"mmap.enabled": True}
+    )
         
     with st.spinner('Parse and chunk documents...'):
         name = 'action_rpg'