Add instructor model

app.py

@@ -290,6 +290,7 @@ elif encoder_model == "Instructor":
     pinecone_index_name = "week13-instructor-xl"
     pinecone_index = pinecone.Index(pinecone_index_name)
     retriever_model = get_instructor_embedding_model()
+    instruction = "Represent the financial question for retrieving supporting documents:"
 
 elif encoder_model == "Hybrid MPNET - SPLADE":
     pinecone.init(
@@ -354,9 +355,14 @@ if document_type == "Single-Document":
         )
 
     else:
-        dense_query_embedding = create_dense_embeddings(
-            query_text, retriever_model
-        )
+        if encoder_model == "Instructor":
+            dense_query_embedding = create_dense_embeddings(
+                query_text, retriever_model, instruction
+            )
+        else:
+            dense_query_embedding = create_dense_embeddings(
+                query_text, retriever_model
+            )
         query_results = query_pinecone(
             dense_query_embedding,
             num_results,
@@ -410,9 +416,14 @@ else:
                 context_group.append((results_list, year, quarter, ticker))
 
         else:
-            dense_query_embedding = create_dense_embeddings(
-                query_text, retriever_model
-            )
+            if encoder_model == "Instructor":
+                dense_query_embedding = create_dense_embeddings(
+                    query_text, retriever_model, instruction
+                )
+            else:
+                dense_query_embedding = create_dense_embeddings(
+                    query_text, retriever_model
+                )
             year_quarter_list = year_quarter_range(
                 start_quarter, start_year, end_quarter, end_year
             )
@@ -494,9 +505,14 @@ else:
                 )
 
         else:
-            dense_query_embedding = create_dense_embeddings(
-                query_text, retriever_model
-            )
+            if encoder_model == "Instructor":
+                dense_query_embedding = create_dense_embeddings(
+                    query_text, retriever_model, instruction
+                )
+            else:
+                dense_query_embedding = create_dense_embeddings(
+                    query_text, retriever_model
+                )
             year_quarter_list = year_quarter_range(
                 start_quarter, start_year, end_quarter, end_year
             )

utils/vector_index.py

@@ -1,8 +1,11 @@
 import torch
 
 
-def create_dense_embeddings(query, model):
-    dense_emb = model.encode([query]).tolist()
+def create_dense_embeddings(query, model, instruction=None):
+    if instruction == None:
+        dense_emb = model.encode([query]).tolist()
+    else:
+        dense_emb = model.encoder([[instruction, query]]).tolist()
     return dense_emb