Added slider for changing number of passages out

app.py

@@ -236,13 +236,15 @@ with block:
         ingest_embed_out = gr.Textbox(label="File/web page preparation progress")
 
     with gr.Tab("Advanced features"):
+        out_passages = gr.Slider(minimum=1, value = 2, maximum=10, step=1, label="Choose number of passages to retrieve from the document. Numbers greater than 2 may lead to increased hallucinations or input text being truncated.")
         with gr.Row():
             model_choice = gr.Radio(label="Choose a chat model", value="Flan Alpaca (small, fast)", choices = ["Flan Alpaca (small, fast)", "Mistral Open Orca (larger, slow)"])
             change_model_button = gr.Button(value="Load model", scale=0)
-        with gr.Accordion("Choose number of model layers to send to GPU (WARNING: please don't modify unless you have a GPU).", open = False):
+        with gr.Accordion("Choose number of model layers to send to GPU (WARNING: please don't modify unless you are sure you have a GPU).", open = False):
             gpu_layer_choice = gr.Slider(label="Choose number of model layers to send to GPU.", value=0, minimum=0, maximum=5, step = 1, visible=True)
             
         load_text = gr.Text(label="Load status")
+        
 
     gr.HTML(
         "<center>This app is based on the models Flan Alpaca and Mistral Open Orca. It powered by Gradio, Transformers, Ctransformers, and Langchain.</a></center>"
@@ -277,14 +279,14 @@ with block:
     # Load in a webpage
 
     # Click/enter to send message action
-    response_click = submit.click(chatf.create_full_prompt, inputs=[message, chat_history_state, current_topic, vectorstore_state, embeddings_state, model_type_state], outputs=[chat_history_state, sources, instruction_prompt_out], queue=False, api_name="retrieval").\
+    response_click = submit.click(chatf.create_full_prompt, inputs=[message, chat_history_state, current_topic, vectorstore_state, embeddings_state, model_type_state, out_passages], outputs=[chat_history_state, sources, instruction_prompt_out], queue=False, api_name="retrieval").\
                 then(chatf.turn_off_interactivity, inputs=[message, chatbot], outputs=[message, chatbot], queue=False).\
                 then(chatf.produce_streaming_answer_chatbot, inputs=[chatbot, instruction_prompt_out, model_type_state], outputs=chatbot)
     response_click.then(chatf.highlight_found_text, [chatbot, sources], [sources]).\
                 then(chatf.add_inputs_answer_to_history,[message, chatbot, current_topic], [chat_history_state, current_topic]).\
                 then(lambda: chatf.restore_interactivity(), None, [message], queue=False)
 
-    response_enter = message.submit(chatf.create_full_prompt, inputs=[message, chat_history_state, current_topic, vectorstore_state, embeddings_state, model_type_state], outputs=[chat_history_state, sources, instruction_prompt_out], queue=False).\
+    response_enter = message.submit(chatf.create_full_prompt, inputs=[message, chat_history_state, current_topic, vectorstore_state, embeddings_state, model_type_state, out_passages], outputs=[chat_history_state, sources, instruction_prompt_out], queue=False).\
                 then(chatf.turn_off_interactivity, inputs=[message, chatbot], outputs=[message, chatbot], queue=False).\
                 then(chatf.produce_streaming_answer_chatbot, [chatbot, instruction_prompt_out, model_type_state], chatbot)    
     response_enter.then(chatf.highlight_found_text, [chatbot, sources], [sources]).\

chatfuncs/chatfuncs.py

@@ -312,7 +312,7 @@ def write_out_metadata_as_string(metadata_in):
     metadata_string = [f"{'  '.join(f'{k}: {v}' for k, v in d.items() if k != 'page_section')}" for d in metadata_in] # ['metadata']
     return metadata_string
 
-def generate_expanded_prompt(inputs: Dict[str, str], instruction_prompt, content_prompt, extracted_memory, vectorstore, embeddings): # , 
+def generate_expanded_prompt(inputs: Dict[str, str], instruction_prompt, content_prompt, extracted_memory, vectorstore, embeddings, out_passages = 2): # , 
         
         question =  inputs["question"]
         chat_history = inputs["chat_history"]
@@ -321,7 +321,7 @@ def generate_expanded_prompt(inputs: Dict[str, str], instruction_prompt, content
         new_question_kworded = adapt_q_from_chat_history(question, chat_history, extracted_memory) # new_question_keywords, 
         
        
-        docs_keep_as_doc, doc_df, docs_keep_out = hybrid_retrieval(new_question_kworded, vectorstore, embeddings, k_val = 25, out_passages = 2,
+        docs_keep_as_doc, doc_df, docs_keep_out = hybrid_retrieval(new_question_kworded, vectorstore, embeddings, k_val = 25, out_passages = out_passages,
                                                                           vec_score_cut_off = 1, vec_weight = 1, bm25_weight = 1, svm_weight = 1)#,
                                                                           #vectorstore=globals()["vectorstore"], embeddings=globals()["embeddings"])
         
@@ -356,7 +356,7 @@ def generate_expanded_prompt(inputs: Dict[str, str], instruction_prompt, content
                 
         return instruction_prompt_out, sources_docs_content_string, new_question_kworded
 
-def create_full_prompt(user_input, history, extracted_memory, vectorstore, embeddings, model_type):
+def create_full_prompt(user_input, history, extracted_memory, vectorstore, embeddings, model_type, out_passages):
     
     if not user_input.strip():
         return history, "", "Respond with 'Please enter a question.' RESPONSE:"
@@ -373,7 +373,7 @@ def create_full_prompt(user_input, history, extracted_memory, vectorstore, embed
     instruction_prompt, content_prompt = base_prompt_templates(model_type=model_type)
     instruction_prompt_out, docs_content_string, new_question_kworded =\
                 generate_expanded_prompt({"question": user_input, "chat_history": history}, #vectorstore,
-                                    instruction_prompt, content_prompt, extracted_memory, vectorstore, embeddings)
+                                    instruction_prompt, content_prompt, extracted_memory, vectorstore, embeddings, out_passages)
     
   
     history.append(user_input)