Update app.py

app.py

@@ -1,5 +1,3 @@
-"Single Thread"
-
 import os
 import multiprocessing
 import concurrent.futures
@@ -10,18 +8,19 @@ from sentence_transformers import SentenceTransformer
 import faiss
 import torch
 import numpy as np
-from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, BitsAndBytesConfig
+from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer, BitsAndBytesConfig
 from datetime import datetime
 import json
 import gradio as gr
-import re 
+import re
+from threading import Thread
 
 class DocumentRetrievalAndGeneration:
     def __init__(self, embedding_model_name, lm_model_id, data_folder):
         self.all_splits = self.load_documents(data_folder)
         self.embeddings = SentenceTransformer(embedding_model_name)
         self.gpu_index = self.create_faiss_index()
-        self.llm = self.initialize_llm(lm_model_id)
+        self.tokenizer, self.model = self.initialize_llm(lm_model_id)
 
     def load_documents(self, folder_path):
         loader = DirectoryLoader(folder_path, loader_cls=TextLoader)
@@ -30,7 +29,7 @@ class DocumentRetrievalAndGeneration:
         all_splits = text_splitter.split_documents(documents)
         print('Length of documents:', len(documents))
         print("LEN of all_splits", len(all_splits))
-        for i in range(5):
+        for i in range(3):
             print(all_splits[i].page_content)
         return all_splits
 
@@ -44,122 +43,99 @@ class DocumentRetrievalAndGeneration:
         return gpu_index
 
     def initialize_llm(self, model_id):
-        bnb_config = BitsAndBytesConfig(
+        quantization_config = BitsAndBytesConfig(
             load_in_4bit=True,
             bnb_4bit_use_double_quant=True,
             bnb_4bit_quant_type="nf4",
             bnb_4bit_compute_dtype=torch.bfloat16
         )
-        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-        model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=bnb_config)
         tokenizer = AutoTokenizer.from_pretrained(model_id)
-        generate_text = pipeline(
-            model=model,
-            tokenizer=tokenizer,
-            return_full_text=True,
-            task='text-generation',
-            temperature=0.6,
-            max_new_tokens=256,
+        model = AutoModelForCausalLM.from_pretrained(
+            model_id,
+            torch_dtype=torch.bfloat16,
+            device_map="auto",
+            quantization_config=quantization_config
         )
-        return generate_text
+        return tokenizer, model
 
-    def generate_response_with_timeout(self, model_inputs):
+    def generate_response_with_timeout(self, input_ids, max_new_tokens=1000):
         try:
-            with concurrent.futures.ThreadPoolExecutor() as executor:
-                future = executor.submit(self.llm.model.generate, model_inputs, max_new_tokens=1000, do_sample=True)
-                generated_ids = future.result(timeout=60)  # Timeout set to 60 seconds
-            return generated_ids
-        except concurrent.futures.TimeoutError:
-            return "Text generation process timed out"
-            raise TimeoutError("Text generation process timed out")
-    
+            streamer = TextIteratorStreamer(self.tokenizer, timeout=60.0, skip_prompt=True, skip_special_tokens=True)
+            generate_kwargs = dict(
+                input_ids=input_ids,
+                max_new_tokens=max_new_tokens,
+                do_sample=True,
+                top_p=1.0,
+                top_k=20,
+                temperature=0.8,
+                repetition_penalty=1.2,
+                eos_token_id=[128001, 128008, 128009],
+                streamer=streamer,
+            )
+            
+            thread = Thread(target=self.model.generate, kwargs=generate_kwargs)
+            thread.start()
+            
+            generated_text = ""
+            for new_text in streamer:
+                generated_text += new_text
+            
+            return generated_text
+        except Exception as e:
+            print(f"Error in generate_response_with_timeout: {str(e)}")
+            return "Text generation process encountered an error"
+
     def query_and_generate_response(self, query):
+        similarityThreshold = 1
         query_embedding = self.embeddings.encode(query, convert_to_tensor=True).cpu().numpy()
-        distances, indices = self.gpu_index.search(np.array([query_embedding]), k=5)
-
+        distances, indices = self.gpu_index.search(np.array([query_embedding]), k=3)
+        print("Distance", distances, "indices", indices)
         content = ""
+        filtered_results = []
         for idx, distance in zip(indices[0], distances[0]):
+            if distance <= similarityThreshold:
+                filtered_results.append(idx)
+            for i in filtered_results:
+                print(self.all_splits[i].page_content)
             content += "-" * 50 + "\n"
             content += self.all_splits[idx].page_content + "\n"
             print("CHUNK", idx)
             print("Distance:", distance)
+            print("indices:", indices)
             print(self.all_splits[idx].page_content)
             print("############################")
-        # for idx in indices[0]:
-        #     content += "-" * 50 + "\n"
-        #     content += self.all_splits[idx].page_content + "\n"
-        #     distance=distances[0][i]
-        #     print("CHUNK", idx)
-        #     print("Distance :",distance)
-        #     print(self.all_splits[idx].page_content)
-        #     print("############################")
-        prompt = f"""<s>
-        You are a knowledgeable assistant with access to a comprehensive database. 
-        I need you to answer my question and provide related information in a specific format.
-        I have provided five relatable json files {content}, choose the most suitable chunks for answering the query
-        RETURN ONLY SOLUTION . IF THEIR IS NO ANSWER RELATABLE IN RETRIEVED CHUNKS , RETURN " NO SOLUTION AVAILABLE"
-        Here's what I need:
-        Include a final answer without additional comments, sign-offs, or extra phrases. Be direct and to the point.
-        content
-        
-        Here's my question:
-        Query:{query}
-        Solution==>
-        
-        
-        Example
-        Query: "How to use IPU1_0 instead of A15_0 to process NDK in TDA2x-EVM",
-        Solution: "To use IPU1_0 instead of A15_0 to process NDK in TDA2x-EVM, you need to modify the configuration file of the NDK application. Specifically, change the processor reference from 'A15_0' to 'IPU1_0'.",
-        
-        Example
-        Query: "Can BQ25896 support I2C interface?",
-        Solution: "Yes, the BQ25896 charger supports the I2C interface for communication."
-
-        Example
-        Query: "Who is the fastest runner in the world",
-        Solution:"NO SOLUTION AVAILABLE"
 
-        Example
-        Query:"What is the price of latest apple MACBOOK "
-        Solution:"NO SOLUTION AVAILABLE"
-
-        RETURN ONLY SOLUTION . IF THEIR IS NO ANSWER RELATABLE IN RETRIEVED CHUNKS , RETURN " NO SOLUTION AVAILABLE"
-        IF THE QUERY AND THE RETRIEVED CHUNKS DO NOT CORRELATE MEANINGFULLY, OR IF THE QUERY IS NOT RELEVANT TO TDA2 OR RELATED TOPICS, THEN "NO SOLUTION AVAILABLE."
-        </s>
-        """
-        # prompt = f"Query: {query}\nSolution: {content}\n"
+        conversation = [
+            {"role": "system", "content": "You are a knowledgeable assistant with access to a comprehensive database."},
+            {"role": "user", "content": f"""
+            I need you to answer my question and provide related information in a specific format.
+            I have provided five relatable json files {content}, choose the most suitable chunks for answering the query.
+            RETURN ONLY SOLUTION. IF THERE IS NO ANSWER RELATABLE IN RETRIEVED CHUNKS, RETURN "NO SOLUTION AVAILABLE".
+            Include a final answer without additional comments, sign-offs, or extra phrases. Be direct and to the point.
+            
+            Here's my question:
+            Query: {query}
+            Solution==>
+            """}
+        ]
 
-        # Encode and prepare inputs
-        messages = [{"role": "user", "content": prompt}]
-        encodeds = self.llm.tokenizer.apply_chat_template(messages, return_tensors="pt")
-        model_inputs = encodeds.to(self.llm.device)
+        input_ids = self.tokenizer.apply_chat_template(conversation, return_tensors="pt").to(self.model.device)
         
-        # Perform inference and measure time
         start_time = datetime.now()
-        generated_ids = self.generate_response_with_timeout(model_inputs)
-        # generated_ids = self.llm.model.generate(model_inputs, max_new_tokens=1000, do_sample=True)
+        generated_response = self.generate_response_with_timeout(input_ids)
         elapsed_time = datetime.now() - start_time
 
-        # Decode and return output
-        decoded = self.llm.tokenizer.batch_decode(generated_ids)
-        generated_response = decoded[0]
-        match1 = re.search(r'\[/INST\](.*?)</s>', generated_response, re.DOTALL)
-        
-        match2 = re.search(r'Solution:(.*?)</s>', generated_response, re.DOTALL | re.IGNORECASE)
-        if match1:
-            solution_text = match1.group(1).strip()
-            print(solution_text)
-            if "Solution:" in solution_text:
-                solution_text = solution_text.split("Solution:", 1)[1].strip()
-        elif match2:
-            solution_text = match2.group(1).strip()
-            print(solution_text)
-            
-        else:
-            solution_text=generated_response
         print("Generated response:", generated_response)
         print("Time elapsed:", elapsed_time)
-        print("Device in use:", self.llm.device)
+        print("Device in use:", self.model.device)
+
+        solution_text = generated_response.strip()
+        if "Solution:" in solution_text:
+            solution_text = solution_text.split("Solution:", 1)[1].strip()
+
+        # Post-processing to remove "assistant" prefix
+        solution_text = re.sub(r'^assistant\s*', '', solution_text, flags=re.IGNORECASE)
+        solution_text = solution_text.strip()
 
         return solution_text, content
 
@@ -168,29 +144,25 @@ class DocumentRetrievalAndGeneration:
         return response
 
 if __name__ == "__main__":
-    # Example usage
     embedding_model_name = 'flax-sentence-embeddings/all_datasets_v3_MiniLM-L12'
-    lm_model_id = "mistralai/Mistral-7B-Instruct-v0.2"
+    lm_model_id = "meta-llama/Meta-Llama-3.1-8B-Instruct"
     data_folder = 'sample_embedding_folder2'
 
     doc_retrieval_gen = DocumentRetrievalAndGeneration(embedding_model_name, lm_model_id, data_folder)
 
-    """Dual Interface"""
-    
     def launch_interface():
         css_code = """
             .gradio-container {
                 background-color: #daccdb;
             }
-            /* Button styling for all buttons */
             button {
-                background-color: #927fc7; /* Default color for all other buttons */
+                background-color: #927fc7;
                 color: black;
                 border: 1px solid black;
                 padding: 10px;
                 margin-right: 10px;
-                font-size: 16px; /* Increase font size */
-                font-weight: bold; /* Make text bold */
+                font-size: 16px;
+                font-weight: bold;
             }
         """
         EXAMPLES = [
@@ -201,20 +173,18 @@ if __name__ == "__main__":
     
         file_path = "ticketNames.txt"
     
-        # Read the file content
         with open(file_path, "r") as file:
             content = file.read()
         ticket_names = json.loads(content)
         dropdown = gr.Dropdown(label="Sample queries", choices=ticket_names)
     
-        # Define Gradio interfaces
         tab1 = gr.Interface(
             fn=doc_retrieval_gen.qa_infer_gradio,
             inputs=[gr.Textbox(label="QUERY", placeholder="Enter your query here")],
             allow_flagging='never',
             examples=EXAMPLES,
             cache_examples=False,
-            outputs=[gr.Textbox(label="SOLUTION"), gr.Textbox(label="RELATED QUERIES")],
+            outputs=[gr.Textbox(label="RESPONSE"), gr.Textbox(label="RELATED QUERIES")],
             css=css_code
         )
         tab2 = gr.Interface(
@@ -225,7 +195,6 @@ if __name__ == "__main__":
             css=css_code
         )
     
-        # Combine interfaces into a tabbed interface
         gr.TabbedInterface(
             [tab1, tab2],
             ["Textbox Input", "FAQs"],
@@ -233,53 +202,4 @@ if __name__ == "__main__":
             css=css_code
         ).launch(debug=True)
     
-    # Launch the interface
-    launch_interface()
-
-
-
-    """Single Interface"""
-    # def launch_interface():
-    #     css_code = """
-    #         .gradio-container {
-    #             background-color: #daccdb;
-    #         }
-    #         /* Button styling for all buttons */
-    #         button {
-    #             background-color: #927fc7; /* Default color for all other buttons */
-    #             color: black;
-    #             border: 1px solid black;
-    #             padding: 10px;
-    #             margin-right: 10px;
-    #             font-size: 16px; /* Increase font size */
-    #             font-weight: bold; /* Make text bold */
-    #         }
-    #         """
-    #     EXAMPLES = ["On which devices can the VIP and CSI2 modules operate simultaneously? ", 
-    #                 "I'm using Code Composer Studio 5.4.0.00091 and enabled FPv4SPD16 floating point support for CortexM4 in TDA2. However, after building the project, the .asm file shows --float_support=vfplib instead of FPv4SPD16. Why is this happening?", 
-    #                 "Could you clarify the maximum number of cameras that can be connected simultaneously to the video input ports on the TDA2x SoC, considering it supports up to 10 multiplexed input ports and includes 3 dedicated video input modules?"]
-        
-    #     file_path = "ticketNames.txt"
-
-    #     # Read the file content
-    #     with open(file_path, "r") as file:
-    #         content = file.read()
-    #     ticket_names = json.loads(content)
-    #     dropdown = gr.Dropdown(label="Sample queries", choices=ticket_names)
-        
-    #     # Define Gradio interface
-    #     interface = gr.Interface(
-    #         fn=doc_retrieval_gen.qa_infer_gradio,
-    #         inputs=[gr.Textbox(label="QUERY", placeholder="Enter your query here")],
-    #         allow_flagging='never',
-    #         examples=EXAMPLES,
-    #         cache_examples=False,
-    #         outputs=[gr.Textbox(label="SOLUTION"), gr.Textbox(label="RELATED QUERIES")],
-    #         css=css_code
-    #     )
-
-    #     # Launch Gradio interface
-    #     interface.launch(debug=True)
-
-    # # Launch the interface
-    # launch_interface()
+    launch_interface()