Update app.py

- Modification du prompt caché
- Modification des paramètres de chunk
-Ajout de l'argument similarity_to_k dans le query engine

app.py

@@ -41,22 +41,19 @@ def construct_index(doc):
 
   ## Define the prompt helper
     # Set maximum input size
-  max_input_size = 400
+  max_input_size = 32000
 
     # Set number of output tokens
-  num_output = 400 # About 300 words
+  num_output = 600 # About 300 words
 
     #Set the chunk size limit 
   chunk_size_limit = 600 # About 450 words ~ 1 page
 
-    # Set maximum chunk overlap
-  max_chunk_overlap = 1
-
     # Set chunk overlap ratio 
   chunk_overlap_ratio = 0.2
 
     # Define prompt helper
-  prompt_helper = PromptHelper(max_input_size, num_output, max_chunk_overlap, chunk_size_limit, chunk_overlap_ratio)
+  prompt_helper = PromptHelper(max_input_size, num_output, chunk_size_limit, chunk_overlap_ratio)
 
   ## Define the LLM predictor 
   llm_predictor = LLMPredictor(llm=ChatOpenAI(temperature=0.4, model_name="gpt-4-32k", max_tokens=num_output))
@@ -114,15 +111,17 @@ def extract_name(file):
         
 def ask_ai_upload(doc, question):
   file_name = extract_name(doc)
+    
   try:
       storage_context = StorageContext.from_defaults(persist_dir=f'gpt/storage_demo/{file_name}', fs=fs)
       # Load index
       index = load_index_from_storage(storage_context)
+      
   except:
+      # Construct index
       text = extract_text(doc)
       index = construct_index([Document(text)])
     
-      
       # Save index to Azure blob storage
       index.storage_context.persist(f'gpt/storage_demo/{file_name}', fs=fs)
     
@@ -133,9 +132,8 @@ def ask_ai_upload(doc, question):
       index = load_index_from_storage(storage_context)
   
   # Define the query & the querying method
-  query_engine = index.as_query_engine(optimizer=SentenceEmbeddingOptimizer(percentile_cutoff=0.8)) 
-  query = 'Your task is to  answer a question on the report loaded and give insights to an investment team in Infrastructure. Make your response as clear and precise as possible. The question is:' + str(question)
-  response = query_engine.query(query)
+  query_engine = index.as_query_engine(optimizer=SentenceEmbeddingOptimizer(percentile_cutoff=0.8), similarity_top_k=7) 
+  query = 'Answer the question truthfully based on the text provided. Include verbatim quote and after the quote write a step by step explanation. Use bullet points. Provide an answer as detailed and precise as possible. The task is:' + str(question)  response = query_engine.query(query)
 
   return response.response
 
@@ -158,9 +156,8 @@ def ask_ai_choose(doc, question):
   index = load_index_from_storage(storage_context)
   
   # Define the query & the querying method
-  query_engine = index.as_query_engine(optimizer=SentenceEmbeddingOptimizer(percentile_cutoff=0.8)) 
-  query = 'Your task is to  answer a question on the report loaded and give insights to an investment team in Infrastructure. Make your response as clear and precise as possible. The question is:' + str(question)
-  response = query_engine.query(query)
+  query_engine = index.as_query_engine(optimizer=SentenceEmbeddingOptimizer(percentile_cutoff=0.8), similarity_top_k=7) 
+  query = 'Answer the question truthfully based on the text provided. Include verbatim quote and after the quote write a step by step explanation. Use bullet points. Provide an answer as detailed and precise as possible. The task is:' + str(question)  response = query_engine.query(query)
 
   return response.response