Update app.py

app.py

@@ -20,6 +20,13 @@ import torch
 import tqdm 
 import accelerate
 
+#set parameters
+slider_chunk_size = 4096
+slider_chunk_overlap = 256
+slider_temperature = 0.1
+slider_maxtokens = 2048
+slider_topk = 3
+llm_model = "mistralai/Mistral-7B-Instruct-v0.2"
 
 
 # default_persist_directory = './chroma_HF/'
@@ -73,62 +80,12 @@ def load_db():
 
 # Initialize langchain LLM chain
 def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
-    progress(0.1, desc="Initializing HF tokenizer...")
-    # HuggingFacePipeline uses local model
-    # Note: it will download model locally...
-    # tokenizer=AutoTokenizer.from_pretrained(llm_model)
-    # progress(0.5, desc="Initializing HF pipeline...")
-    # pipeline=transformers.pipeline(
-    #     "text-generation",
-    #     model=llm_model,
-    #     tokenizer=tokenizer,
-    #     torch_dtype=torch.bfloat16,
-    #     trust_remote_code=True,
-    #     device_map="auto",
-    #     # max_length=1024,
-    #     max_new_tokens=max_tokens,
-    #     do_sample=True,
-    #     top_k=top_k,
-    #     num_return_sequences=1,
-    #     eos_token_id=tokenizer.eos_token_id
-    #     )
-    # llm = HuggingFacePipeline(pipeline=pipeline, model_kwargs={'temperature': temperature})
-    
-    # HuggingFaceHub uses HF inference endpoints
-    progress(0.5, desc="Initializing HF Hub...")
-    # Use of trust_remote_code as model_kwargs
-    # Warning: langchain issue
-    # URL: https://github.com/langchain-ai/langchain/issues/6080
-    if llm_model == "mistralai/Mixtral-8x7B-Instruct-v0.1":
-        llm = HuggingFaceHub(
-            repo_id=llm_model, 
-            model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k, "load_in_8bit": True}
-        )
-    elif llm_model == "microsoft/phi-2":
-        raise gr.Error("phi-2 model requires 'trust_remote_code=True', currently not supported by langchain HuggingFaceHub...")
-        llm = HuggingFaceHub(
-            repo_id=llm_model, 
-            model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k, "trust_remote_code": True, "torch_dtype": "auto"}
-        )
-    elif llm_model == "TinyLlama/TinyLlama-1.1B-Chat-v1.0":
-        llm = HuggingFaceHub(
-            repo_id=llm_model, 
-            model_kwargs={"temperature": temperature, "max_new_tokens": 250, "top_k": top_k}
-        )
-    elif llm_model == "meta-llama/Llama-2-7b-chat-hf":
-        raise gr.Error("Llama-2-7b-chat-hf model requires a Pro subscription...")
-        llm = HuggingFaceHub(
-            repo_id=llm_model, 
-            model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k}
-        )
-    else:
-        llm = HuggingFaceHub(
-            repo_id=llm_model, 
-            # model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k, "trust_remote_code": True, "torch_dtype": "auto"}
-            model_kwargs={"temperature": temperature, "max_new_tokens": max_tokens, "top_k": top_k}
-        )
+    llm = HuggingFaceHub(repo_id=llm_model, model_kwargs={"temperature": 
+                                                          temperature, "max_new_tokens": 
+                                                          max_tokens, "top_k": top_k, 
+                                                          "load_in_8bit": True})
+
     
-    progress(0.75, desc="Defining buffer memory...")
     memory = ConversationBufferMemory(
         memory_key="chat_history",
         output_key='answer',
@@ -136,7 +93,6 @@ def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, pr
     )
     # retriever=vector_db.as_retriever(search_type="similarity", search_kwargs={'k': 3})
     retriever=vector_db.as_retriever()
-    progress(0.8, desc="Defining retrieval chain...")
     qa_chain = ConversationalRetrievalChain.from_llm(
         llm,
         retriever=retriever,
@@ -147,7 +103,6 @@ def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, pr
         #return_generated_question=False,
         verbose=False,
     )
-    progress(0.9, desc="Done!")
     return qa_chain
 
 
@@ -238,64 +193,28 @@ def demo():
         vector_db = gr.State()
         qa_chain = gr.State()
         collection_name = gr.State()
-        
-        gr.Markdown(
-        """<center><h2>PDF-based chatbot (powered by LangChain and open-source LLMs)</center></h2>
-        <h3>Ask any questions about your PDF documents, along with follow-ups</h3>
-        <b>Note:</b> This AI assistant performs retrieval-augmented generation from your PDF documents. \
-        When generating answers, it takes past questions into account (via conversational memory), and includes document references for clarity purposes.</i>
-        <br><b>Warning:</b> This space uses the free CPU Basic hardware from Hugging Face. Some steps and LLM models used below (free inference endpoints) can take some time to generate an output.<br>
-        """)
-        with gr.Tab("Step 1 - Document pre-processing"):
-            with gr.Row():
-                document = gr.Files(value = ['/home/user/app/pdfs/Annual-Report-2022-2023-English_1.pdf'],visible=False,
-                                    height=100, file_count="multiple", file_types=["pdf"], label="Upload your PDF documents (single or multiple)")
-                # upload_btn = gr.UploadButton("Loading document...", height=100, file_count="multiple", file_types=["pdf"], scale=1)
-            with gr.Row():
-                db_btn = gr.Radio(["ChromaDB"], label="Vector database type", value = "ChromaDB", type="index", info="Choose your vector database")
-            with gr.Accordion("Advanced options - Document text splitter", open=False):
-                with gr.Row():
-                    slider_chunk_size = gr.Slider(minimum = 100, maximum = 1000, value=600, step=20, label="Chunk size", info="Chunk size", interactive=True)
-                with gr.Row():
-                    slider_chunk_overlap = gr.Slider(minimum = 10, maximum = 200, value=40, step=10, label="Chunk overlap", info="Chunk overlap", interactive=True)
-            with gr.Row():
-                db_progress = gr.Textbox(label="Vector database initialization", value="None")
-            with gr.Row():
-                db_btn = gr.Button("Generate vector database...")
-            
-        with gr.Tab("Step 2 - QA chain initialization"):
-            with gr.Row():
-                llm_btn = gr.Radio(list_llm_simple, \
-                    label="LLM models", value = list_llm_simple[0], type="index", info="Choose your LLM model")
-            with gr.Accordion("Advanced options - LLM model", open=False):
-                with gr.Row():
-                    slider_temperature = gr.Slider(minimum = 0.0, maximum = 1.0, value=0.7, step=0.1, label="Temperature", info="Model temperature", interactive=True)
-                with gr.Row():
-                    slider_maxtokens = gr.Slider(minimum = 224, maximum = 4096, value=1024, step=32, label="Max Tokens", info="Model max tokens", interactive=True)
-                with gr.Row():
-                    slider_topk = gr.Slider(minimum = 1, maximum = 10, value=3, step=1, label="top-k samples", info="Model top-k samples", interactive=True)
-            with gr.Row():
-                llm_progress = gr.Textbox(value="None",label="QA chain initialization")
-            with gr.Row():
-                qachain_btn = gr.Button("Initialize question-answering chain...")
 
-        with gr.Tab("Step 3 - Conversation with chatbot"):
-            chatbot = gr.Chatbot(height=300)
-            with gr.Accordion("Advanced - Document references", open=False):
-                with gr.Row():
-                    doc_source1 = gr.Textbox(label="Reference 1", lines=2, container=True, scale=20)
-                    source1_page = gr.Number(label="Page", scale=1)
-                with gr.Row():
-                    doc_source2 = gr.Textbox(label="Reference 2", lines=2, container=True, scale=20)
-                    source2_page = gr.Number(label="Page", scale=1)
-                with gr.Row():
-                    doc_source3 = gr.Textbox(label="Reference 3", lines=2, container=True, scale=20)
-                    source3_page = gr.Number(label="Page", scale=1)
+        document = gr.Files(value = ['/home/user/app/pdfs/Annual-Report-2022-2023-English_1.pdf'],visible=False,
+                                height=100, file_count="multiple", file_types=["pdf"], label="Upload your PDF documents (single or multiple)")
+        chatbot = gr.Chatbot(height=300)
+        db_btn = gr.Radio(["ChromaDB"], label="Vector database type", value = "ChromaDB", type="index", info="Choose your vector database", visible=False)
+        with gr.Accordion("Advanced - Document references", open=False):
+            with gr.Row():
+                doc_source1 = gr.Textbox(label="Reference 1", lines=2, container=True, scale=20)
+                source1_page = gr.Number(label="Page", scale=1)
             with gr.Row():
-                msg = gr.Textbox(placeholder="Type message", container=True)
+                doc_source2 = gr.Textbox(label="Reference 2", lines=2, container=True, scale=20)
+                source2_page = gr.Number(label="Page", scale=1)
             with gr.Row():
-                submit_btn = gr.Button("Submit")
-                clear_btn = gr.ClearButton([msg, chatbot])
+                doc_source3 = gr.Textbox(label="Reference 3", lines=2, container=True, scale=20)
+                source3_page = gr.Number(label="Page", scale=1)
+        with gr.Row():
+            msg = gr.Textbox(placeholder="Type message", container=True)
+        with gr.Row():
+            db_btn = gr.Button("Generate vector database...")
+            qachain_btn = gr.Button("Initialize question-answering chain...")
+            submit_btn = gr.Button("Submit")
+            clear_btn = gr.ClearButton([msg, chatbot])
             
         # Preprocessing events
         #upload_btn.upload(upload_file, inputs=[upload_btn], outputs=[document])
@@ -303,7 +222,7 @@ def demo():
             inputs=[document, slider_chunk_size, slider_chunk_overlap], \
             outputs=[vector_db, collection_name, db_progress])
         qachain_btn.click(initialize_LLM, \
-            inputs=[llm_btn, slider_temperature, slider_maxtokens, slider_topk, vector_db], \
+            inputs=[llm_model, slider_temperature, slider_maxtokens, slider_topk, vector_db], \
             outputs=[qa_chain, llm_progress]).then(lambda:[None,"",0,"",0,"",0], \
             inputs=None, \
             outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \