all done

app.py

@@ -22,15 +22,6 @@ import tqdm
 import accelerate
 import re
 
-
-import torch
-from sacrebleu import corpus_bleu
-from rouge_score import rouge_scorer
-from bert_score import score
-from transformers import GPT2LMHeadModel, GPT2Tokenizer, pipeline
-import nltk
-from nltk.util import ngrams
-
 api_key = os.getenv('API_KEY')
 
 
@@ -87,6 +78,25 @@ 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...")
@@ -237,138 +247,32 @@ def format_chat_history(message, chat_history):
         formatted_chat_history.append(f"User: {user_message}")
         formatted_chat_history.append(f"Assistant: {bot_message}")
     return formatted_chat_history
+    
 
-#----------------------------------------------------------------------------------
-def load_gpt2_model():
-    model = GPT2LMHeadModel.from_pretrained('gpt2')
-    tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
-    return model, tokenizer
-
-gpt2_model, gpt2_tokenizer = load_gpt2_model()
-bias_pipeline = pipeline("zero-shot-classification", model="Hate-speech-CNERG/dehatebert-mono-english")
-
-def evaluate_bleu_rouge(candidates, references):
-    bleu_score = corpus_bleu(candidates, [references]).score
-    scorer = rouge_scorer.RougeScorer(['rouge1', 'rouge2', 'rougeL'], use_stemmer=True)
-    rouge_scores = [scorer.score(ref, cand) for ref, cand in zip(references, candidates)]
-    rouge1 = sum([score['rouge1'].fmeasure for score in rouge_scores]) / len(rouge_scores)
-    return bleu_score, rouge1
-
-def evaluate_bert_score(candidates, references):
-    P, R, F1 = score(candidates, references, lang="en", model_type='bert-base-multilingual-cased')
-    return P.mean().item(), R.mean().item(), F1.mean().item()
-
-def evaluate_perplexity(text, model, tokenizer):
-    encodings = tokenizer(text, return_tensors='pt')
-    max_length = model.config.n_positions
-    stride = 512
-    lls = []
-    for i in range(0, encodings.input_ids.size(1), stride):
-        begin_loc = max(i + stride - max_length, 0)
-        end_loc = min(i + stride, encodings.input_ids.size(1))
-        trg_len = end_loc - i
-        input_ids = encodings.input_ids[:, begin_loc:end_loc]
-        target_ids = input_ids.clone()
-        target_ids[:, :-trg_len] = -100
-        with torch.no_grad():
-            outputs = model(input_ids, labels=target_ids)
-            log_likelihood = outputs[0] * trg_len
-        lls.append(log_likelihood)
-    ppl = torch.exp(torch.stack(lls).sum() / end_loc)
-    return ppl.item()
-
-def evaluate_diversity(texts):
-    all_tokens = [tok for text in texts for tok in text.split()]
-    unique_bigrams = set(ngrams(all_tokens, 2))
-    diversity_score = len(unique_bigrams) / len(all_tokens) if all_tokens else 0
-    return diversity_score
-
-def evaluate_racial_bias(text, pipeline):
-    results = pipeline([text], candidate_labels=["hate speech", "not hate speech"])
-    bias_score = results[0]['scores'][results[0]['labels'].index('hate speech')]
-    return bias_score
-
-def evaluate_all(question, response, reference, gpt2_model, gpt2_tokenizer, bias_pipeline):
-    candidates = [response]
-    references = [reference]
-    bleu, rouge1 = evaluate_bleu_rouge(candidates, references)
-    bert_p, bert_r, bert_f1 = evaluate_bert_score(candidates, references)
-    perplexity = evaluate_perplexity(response, gpt2_model, gpt2_tokenizer)
-    diversity = evaluate_diversity(candidates)
-    racial_bias = evaluate_racial_bias(response, bias_pipeline)
-    return {
-        "BLEU": bleu,
-        "ROUGE-1": rouge1,
-        "BERT P": bert_p,
-        "BERT R": bert_r,
-        "BERT F1": bert_f1,
-        "Perplexity": perplexity,
-        "Diversity": diversity,
-        "Racial Bias": racial_bias
-    }
-
-#---------------------------------------------------------------------------------
-
-def display_metrics(metrics):
-    result = ""
-    for k, v in metrics.items():
-        if k == 'BLEU':
-            result += f"BLEU measures the overlap between the generated output and reference text based on n-grams. Higher scores indicate better match. Score obtained: {v}\n\n"
-        elif k == "ROUGE-1":
-            result += f"ROUGE-1 measures the overlap of unigrams between the generated output and reference text. Higher scores indicate better match. Score obtained: {v}\n\n"
-        elif k == 'BERT P':
-            result += "BERTScore evaluates the semantic similarity between the generated output and reference text using BERT embeddings.\n\n"
-            result += f"**BERT Precision**: {metrics['BERT P']}\n"
-            result += f"**BERT Recall**: {metrics['BERT R']}\n"
-            result += f"**BERT F1 Score**: {metrics['BERT F1']}\n\n"
-        elif k == 'Perplexity':
-            result += f"Perplexity measures how well a language model predicts the text. Lower values indicate better fluency and coherence. Score obtained: {v}\n\n"
-        elif k == 'Diversity':
-            result += f"Diversity measures the uniqueness of bigrams in the generated output. Higher values indicate more diverse and varied output. Score obtained: {v}\n\n"
-        elif k == 'Racial Bias':
-            result += f"Racial Bias score indicates the presence of biased language in the generated output. Higher scores indicate more bias. Score obtained: {v}\n\n"
-    return result
-#---------------------------------------------------------------------------------------------------------------------------------------------------
-
-
-
-
-
-
-def conversation(qa_chain, message, history, gpt2_model, gpt2_tokenizer, bias_pipeline):
+def conversation(qa_chain, message, history):
     formatted_chat_history = format_chat_history(message, history)
-    question_by_user = message
-    
+    #print("formatted_chat_history",formatted_chat_history)
+   
+    # Generate response using QA chain
     response = qa_chain({"question": message, "chat_history": formatted_chat_history})
     response_answer = response["answer"]
-    answer_of_question = response["answer"]
     if response_answer.find("Helpful Answer:") != -1:
         response_answer = response_answer.split("Helpful Answer:")[-1]
     response_sources = response["source_documents"]
-    context = " ".join([d.page_content for d in response_sources])
-    
     response_source1 = response_sources[0].page_content.strip()
     response_source2 = response_sources[1].page_content.strip()
     response_source3 = response_sources[2].page_content.strip()
-
+    # Langchain sources are zero-based
     response_source1_page = response_sources[0].metadata["page"] + 1
     response_source2_page = response_sources[1].metadata["page"] + 1
     response_source3_page = response_sources[2].metadata["page"] + 1
-
-    new_history = history + [(message, response_answer)]
+    # print ('chat response: ', response_answer)
+    # print('DB source', response_sources)
     
-    # Evaluate the metrics
-    metrics = evaluate_all(question_by_user, answer_of_question, context,gpt2_model, gpt2_tokenizer, bias_pipeline)
-    evaluation_metrics = display_metrics(metrics)
-
-    return (qa_chain, gr.update(value=""), new_history, response_source1, response_source1_page,
-            response_source2, response_source2_page, response_source3, response_source3_page,
-            evaluation_metrics)
-
-
-# def interact(qa_chain, message, history):
-#     return conversation(qa_chain, message, history, evaluator)
-            
+    # Append user message and response to chat history
+    new_history = history + [(message, response_answer)]
+    # return gr.update(value=""), new_history, response_sources[0], response_sources[1] 
+    return qa_chain, gr.update(value=""), new_history, response_source1, response_source1_page, response_source2, response_source2_page, response_source3, response_source3_page
     
 
 def upload_file(file_obj):
@@ -380,21 +284,19 @@ def upload_file(file_obj):
     # initialize_database(file_path, progress)
     return list_file_path
 
-####################################
 
 def demo():
     with gr.Blocks(theme="base") as demo:
         vector_db = gr.State()
         qa_chain = gr.State()
         collection_name = gr.State()
-        history = gr.State([])  # Initialize history as an empty list
         
         gr.Markdown(
         """<center><h2>PDF-based chatbot</center></h2>
         <h3>Ask any questions about your PDF documents</h3>""")
         gr.Markdown(
         """<b>Note:</b> This AI assistant, using Langchain and open-source LLMs, performs retrieval-augmented generation (RAG) from your PDF documents. \
-        The user interface explicitly shows multiple steps to help understand the RAG workflow. 
+        The user interface explicitely shows multiple steps to help understand the RAG workflow. 
         This chatbot takes past questions into account when generating answers (via conversational memory), and includes document references for clarity purposes.<br>
         <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 a reply.
         """)
@@ -402,15 +304,16 @@ def demo():
         with gr.Tab("Step 1 - Upload PDF"):
             with gr.Row():
                 document = gr.Files(height=100, file_count="multiple", file_types=["pdf"], interactive=True, 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.Tab("Step 2 - Process document"):
             with gr.Row():
-                db_btn = gr.Radio(["ChromaDB"], label="Vector database type", value="ChromaDB", type="index", info="Choose your vector database")
+                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)
+                    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)
+                    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():
@@ -418,16 +321,17 @@ def demo():
             
         with gr.Tab("Step 3 - Initialize QA chain"):
             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")
+                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=0.01, maximum=1.0, value=0.7, step=0.1, label="Temperature", info="Model temperature", interactive=True)
+                    slider_temperature = gr.Slider(minimum = 0.01, 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)
+                    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)
+                    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")
+                llm_progress = gr.Textbox(value="None",label="QA chain initialization")
             with gr.Row():
                 qachain_btn = gr.Button("Initialize Question Answering chain")
 
@@ -448,153 +352,34 @@ def demo():
             with gr.Row():
                 submit_btn = gr.Button("Submit message")
                 clear_btn = gr.ClearButton([msg, chatbot], value="Clear conversation")
-            with gr.Row("Metrics"):
-                metrics_output = gr.Textbox(lines=10, label="Evaluation Metrics")
-
+            
         # Preprocessing events
-        db_btn.click(initialize_database, 
-                     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], 
-                          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], 
-                                                                 queue=False)
+        #upload_btn.upload(upload_file, inputs=[upload_btn], outputs=[document])
+        db_btn.click(initialize_database, \
+            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], \
+            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], \
+            queue=False)
 
         # Chatbot events
         msg.submit(conversation, \
-            inputs=[qa_chain, msg, chatbot, gpt2_model, gpt2_tokenizer, bias_pipeline], \
-            outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page,metrics_output], \
+            inputs=[qa_chain, msg, chatbot], \
+            outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
+            queue=False)
+        submit_btn.click(conversation, \
+            inputs=[qa_chain, msg, chatbot], \
+            outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
+            queue=False)
+        clear_btn.click(lambda:[None,"",0,"",0,"",0], \
+            inputs=None, \
+            outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
             queue=False)
-
-        submit_btn.click(conversation, 
-                         inputs=[qa_chain, msg, history, gpt2_model, gpt2_tokenizer, bias_pipeline], 
-                         outputs=[qa_chain, chatbot, history, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page, metrics_output])
-
-        clear_btn.click(lambda: [None, "", 0, "", 0, "", 0], 
-                        inputs=None, 
-                        outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], 
-                        queue=False)
-
     demo.queue().launch(debug=True)
 
-if __name__ == "__main__":
-    demo()
-
-
-
-
 
-###################################
-# def demo():
-#     with gr.Blocks(theme="base") as demo:
-#         vector_db = gr.State()
-#         qa_chain = gr.State()
-#         collection_name = gr.State()
-#         history = gr.State()
-        
-#         gr.Markdown(
-#         """<center><h2>PDF-based chatbot</center></h2>
-#         <h3>Ask any questions about your PDF documents</h3>""")
-#         gr.Markdown(
-#         """<b>Note:</b> This AI assistant, using Langchain and open-source LLMs, performs retrieval-augmented generation (RAG) from your PDF documents. \
-#         The user interface explicitely shows multiple steps to help understand the RAG workflow. 
-#         This chatbot takes past questions into account when generating answers (via conversational memory), and includes document references for clarity purposes.<br>
-#         <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 a reply.
-#         """)
-        
-#         with gr.Tab("Step 1 - Upload PDF"):
-#             with gr.Row():
-#                 document = gr.Files(height=100, file_count="multiple", file_types=["pdf"], interactive=True, 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.Tab("Step 2 - Process document"):
-#             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 3 - Initialize QA chain"):
-        #     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 = 0.01, 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 4 - 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)
-        #     with gr.Row():
-        #         msg = gr.Textbox(placeholder="Type message (e.g. 'What is this document about?')", container=True)
-        #     with gr.Row():
-        #         submit_btn = gr.Button("Submit message")
-        #         clear_btn = gr.ClearButton([msg, chatbot], value="Clear conversation")
-        #     with gr.Row("Metrics"):
-        #         metrics_output = gr.Textbox(lines=10, label="Evaluation Metrics")
-
-            
-        # # Preprocessing events
-        # #upload_btn.upload(upload_file, inputs=[upload_btn], outputs=[document])
-        # db_btn.click(initialize_database, \
-        #     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], \
-        #     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], \
-        #     queue=False)
-
-        # Chatbot events
-    #     msg.submit(interact, inputs=[gr.State(),qa_chain, msg, history], outputs=[
-    #     gr.State(), chatbot, history, response_source1, response_source1_page,
-    #     response_source2, response_source2_page, response_source3, response_source3_page,
-    #     None, None, None, metrics_output
-    # ],queue=False)
-
-
-    #     submit_btn.click(conversation, \
-    #         inputs=[qa_chain, msg, chatbot], \
-    #         outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
-    #         queue=False)
-    #     clear_btn.click(lambda:[None,"",0,"",0,"",0], \
-    #         inputs=None, \
-    #         outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], \
-    #         queue=False)
-
-
-        
-
-
-    # demo.queue().launch(debug=True)
-
-
-# if __name__ == "__main__":
-#     demo()
+if __name__ == "__main__":
+    demo()