Update app.py

app.py

@@ -8,7 +8,7 @@ from langchain.text_splitter import (
     CharacterTextSplitter,
     TokenTextSplitter
 )
-from langchain_community.vectorstores import FAISS, Chroma
+from langchain_community.vectorstores import FAISS, Chroma, Qdrant
 from langchain_community.document_loaders import PyPDFLoader
 from langchain.chains import ConversationalRetrievalChain
 from langchain_community.embeddings import HuggingFaceEmbeddings 
@@ -25,12 +25,19 @@ from ragas.metrics import (
 )
 import pandas as pd
 
-# Constants and configurations
+# Constants and setup
+list_llm = ["meta-llama/Meta-Llama-3-8B-Instruct", "mistralai/Mistral-7B-Instruct-v0.2"]
+list_llm_simple = [os.path.basename(llm) for llm in list_llm]
+api_token = os.getenv("HF_TOKEN")
+
 CHUNK_SIZES = {
     "small": {"recursive": 512, "fixed": 512, "token": 256},
     "medium": {"recursive": 1024, "fixed": 1024, "token": 512}
 }
 
+# Initialize sentence transformer for evaluation
+sentence_model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
+
 class RAGEvaluator:
     def __init__(self):
         self.datasets = {
@@ -51,7 +58,7 @@ class RAGEvaluator:
                     "context": sample["context"]
                 }
                 for sample in samples
-                if sample["answers"]["text"]  # Filter out samples without answers
+                if sample["answers"]["text"]
             ]
         elif dataset_name == "msmarco":
             dataset = load_dataset("ms_marco", "v2.1", split="train")
@@ -63,16 +70,12 @@ class RAGEvaluator:
                     "context": sample["passages"]["passage_text"][0]
                 }
                 for sample in samples
-                if sample["answers"]  # Filter out samples without answers
+                if sample["answers"]
             ]
         self.current_dataset = dataset_name
         return self.test_samples
-    
-    def evaluate_configuration(self, 
-                             vector_db,
-                             qa_chain,
-                             splitting_strategy: str,
-                             chunk_size: str) -> Dict:
+
+    def evaluate_configuration(self, vector_db, qa_chain, splitting_strategy: str, chunk_size: str) -> Dict:
         if not self.test_samples:
             return {"error": "No dataset loaded"}
             
@@ -90,21 +93,9 @@ class RAGEvaluator:
                 "ground_truths": [sample["ground_truth"]]
             })
         
-        # Convert to RAGAS dataset format
         eval_dataset = Dataset.from_list(results)
-        
-        # Calculate RAGAS metrics
-        metrics = [
-            ContextRecall(),
-            AnswerRelevancy(),
-            Faithfulness(),
-            ContextPrecision()
-        ]
-        
-        scores = evaluate(
-            eval_dataset,
-            metrics=metrics
-        )
+        metrics = [ContextRecall(), AnswerRelevancy(), Faithfulness(), ContextPrecision()]
+        scores = evaluate(eval_dataset, metrics=metrics)
         
         return {
             "configuration": f"{splitting_strategy}_{chunk_size}",
@@ -120,6 +111,102 @@ class RAGEvaluator:
             ]))
         }
 
+# Text splitting and database functions
+def get_text_splitter(strategy: str, chunk_size: int = 1024, chunk_overlap: int = 64):
+    splitters = {
+        "recursive": RecursiveCharacterTextSplitter(
+            chunk_size=chunk_size,
+            chunk_overlap=chunk_overlap
+        ),
+        "fixed": CharacterTextSplitter(
+            chunk_size=chunk_size,
+            chunk_overlap=chunk_overlap
+        ),
+        "token": TokenTextSplitter(
+            chunk_size=chunk_size,
+            chunk_overlap=chunk_overlap
+        )
+    }
+    return splitters.get(strategy)
+
+def load_doc(list_file_path: List[str], splitting_strategy: str, chunk_size: str):
+    chunk_size_value = CHUNK_SIZES[chunk_size][splitting_strategy]
+    loaders = [PyPDFLoader(x) for x in list_file_path]
+    pages = []
+    for loader in loaders:
+        pages.extend(loader.load())
+    
+    text_splitter = get_text_splitter(splitting_strategy, chunk_size_value)
+    doc_splits = text_splitter.split_documents(pages)
+    return doc_splits
+
+def create_db(splits, db_choice: str = "faiss"):
+    embeddings = HuggingFaceEmbeddings()
+    db_creators = {
+        "faiss": lambda: FAISS.from_documents(splits, embeddings),
+        "chroma": lambda: Chroma.from_documents(splits, embeddings),
+        "qdrant": lambda: Qdrant.from_documents(
+            splits,
+            embeddings,
+            location=":memory:",
+            collection_name="pdf_docs"
+        )
+    }
+    return db_creators[db_choice]()
+
+def initialize_database(list_file_obj, splitting_strategy, chunk_size, db_choice, progress=gr.Progress()):
+    list_file_path = [x.name for x in list_file_obj if x is not None]
+    doc_splits = load_doc(list_file_path, splitting_strategy, chunk_size)
+    vector_db = create_db(doc_splits, db_choice)
+    return vector_db, f"Database created using {splitting_strategy} splitting and {db_choice} vector database!"
+
+def initialize_llmchain(llm_choice, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
+    llm_model = list_llm[llm_choice]
+    
+    llm = HuggingFaceEndpoint(
+        repo_id=llm_model,
+        huggingfacehub_api_token=api_token,
+        temperature=temperature,
+        max_new_tokens=max_tokens,
+        top_k=top_k
+    )
+    
+    memory = ConversationBufferMemory(
+        memory_key="chat_history",
+        output_key='answer',
+        return_messages=True
+    )
+
+    retriever = vector_db.as_retriever()
+    qa_chain = ConversationalRetrievalChain.from_llm(
+        llm,
+        retriever=retriever,
+        memory=memory,
+        return_source_documents=True
+    )
+    return qa_chain, "LLM initialized successfully!"
+
+def conversation(qa_chain, message, history):
+    response = qa_chain.invoke({
+        "question": message,
+        "chat_history": [(hist[0], hist[1]) for hist in history]
+    })
+    
+    response_answer = response["answer"]
+    if "Helpful Answer:" in response_answer:
+        response_answer = response_answer.split("Helpful Answer:")[-1]
+    
+    sources = response["source_documents"][:3]
+    source_contents = [s.page_content.strip() for s in sources]
+    source_pages = [s.metadata.get("page", 0) + 1 for s in sources]
+    
+    while len(sources) < 3:
+        source_contents.append("")
+        source_pages.append(0)
+    
+    return (qa_chain, gr.update(value=""), history + [(message, response_answer)] + 
+            source_contents + source_pages)
+
 def demo():
     evaluator = RAGEvaluator()
     
@@ -132,12 +219,17 @@ def demo():
         with gr.Tabs():
             # Custom PDF Tab
             with gr.Tab("Custom PDF Chat"):
-                # Your existing UI components here
                 with gr.Row():
                     with gr.Column(scale=86):
                         gr.Markdown("<b>Step 1 - Configure and Initialize RAG Pipeline</b>")
                         with gr.Row():
-                            document = gr.Files(height=300, file_count="multiple", file_types=["pdf"], interactive=True, label="Upload PDF documents")
+                            document = gr.Files(
+                                height=300,
+                                file_count="multiple",
+                                file_types=["pdf"],
+                                interactive=True,
+                                label="Upload PDF documents"
+                            )
                         
                         with gr.Row():
                             splitting_strategy = gr.Radio(
@@ -145,8 +237,8 @@ def demo():
                                 label="Text Splitting Strategy",
                                 value="recursive"
                             )
-                            db_choice = gr.Dropdown(
-                                ["faiss", "chroma"],
+                            db_choice = gr.Radio(
+                                ["faiss", "chroma", "qdrant"],
                                 label="Vector Database",
                                 value="faiss"
                             )
@@ -156,7 +248,79 @@ def demo():
                                 value="medium"
                             )
                         
-                        # Rest of your existing UI components...
+                        with gr.Row():
+                            db_btn = gr.Button("Create vector database")
+                            db_progress = gr.Textbox(
+                                value="Not initialized",
+                                show_label=False
+                            )
+                        
+                        gr.Markdown("<b>Step 2 - Configure LLM</b>")
+                        with gr.Row():
+                            llm_choice = gr.Radio(
+                                list_llm_simple,
+                                label="Available LLMs",
+                                value=list_llm_simple[0],
+                                type="index"
+                            )
+                        
+                        with gr.Row():
+                            with gr.Accordion("LLM Parameters", open=False):
+                                temperature = gr.Slider(
+                                    minimum=0.01,
+                                    maximum=1.0,
+                                    value=0.5,
+                                    step=0.1,
+                                    label="Temperature"
+                                )
+                                max_tokens = gr.Slider(
+                                    minimum=128,
+                                    maximum=4096,
+                                    value=2048,
+                                    step=128,
+                                    label="Max Tokens"
+                                )
+                                top_k = gr.Slider(
+                                    minimum=1,
+                                    maximum=10,
+                                    value=3,
+                                    step=1,
+                                    label="Top K"
+                                )
+                        
+                        with gr.Row():
+                            init_llm_btn = gr.Button("Initialize LLM")
+                            llm_progress = gr.Textbox(
+                                value="Not initialized",
+                                show_label=False
+                            )
+                    
+                    with gr.Column(scale=200):
+                        gr.Markdown("<b>Step 3 - Chat with Documents</b>")
+                        chatbot = gr.Chatbot(height=505)
+                        
+                        with gr.Accordion("Source References", open=False):
+                            with gr.Row():
+                                source1 = gr.Textbox(label="Source 1", lines=2)
+                                page1 = gr.Number(label="Page")
+                            with gr.Row():
+                                source2 = gr.Textbox(label="Source 2", lines=2)
+                                page2 = gr.Number(label="Page")
+                            with gr.Row():
+                                source3 = gr.Textbox(label="Source 3", lines=2)
+                                page3 = gr.Number(label="Page")
+                        
+                        with gr.Row():
+                            msg = gr.Textbox(
+                                placeholder="Ask a question",
+                                show_label=False
+                            )
+                        with gr.Row():
+                            submit_btn = gr.Button("Submit")
+                            clear_btn = gr.ClearButton(
+                                [msg, chatbot],
+                                value="Clear Chat"
+                            )
 
             # Evaluation Tab
             with gr.Tab("RAG Evaluation"):
@@ -188,6 +352,30 @@ def demo():
                     evaluation_results = gr.DataFrame(label="Evaluation Results")
 
         # Event handlers
+        db_btn.click(
+            initialize_database,
+            inputs=[document, splitting_strategy, chunk_size, db_choice],
+            outputs=[vector_db, db_progress]
+        )
+        
+        init_llm_btn.click(
+            initialize_llmchain,
+            inputs=[llm_choice, temperature, max_tokens, top_k, vector_db],
+            outputs=[qa_chain, llm_progress]
+        )
+        
+        msg.submit(
+            conversation,
+            inputs=[qa_chain, msg, chatbot],
+            outputs=[qa_chain, msg, chatbot, source1, page1, source2, page2, source3, page3]
+        )
+        
+        submit_btn.click(
+            conversation,
+            inputs=[qa_chain, msg, chatbot],
+            outputs=[qa_chain, msg, chatbot, source1, page1, source2, page2, source3, page3]
+        )
+
         def load_dataset_handler(dataset_name):
             samples = evaluator.load_dataset(dataset_name)
             return {
@@ -199,7 +387,7 @@ def demo():
         def run_evaluation(dataset_choice, splitting_strategy, chunk_size, vector_db, qa_chain):
             if not evaluator.current_dataset:
                 return pd.DataFrame()
-                
+            
             results = evaluator.evaluate_configuration(
                 vector_db=vector_db,
                 qa_chain=qa_chain,
@@ -207,11 +395,8 @@ def demo():
                 chunk_size=chunk_size
             )
             
-            # Convert results to DataFrame
-            df = pd.DataFrame([results])
-            return df
-        
-        # Connect event handlers
+            return pd.DataFrame([results])
+
         load_dataset_btn.click(
             load_dataset_handler,
             inputs=[dataset_choice],
@@ -229,36 +414,14 @@ def demo():
             ],
             outputs=[evaluation_results]
         )
-        
-        qachain_btn.click(
-            initialize_llmchain,  # Fixed function name here
-            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
-        )
-        
-        msg.submit(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
-        )
-        
-        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 button handlers
         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
+            outputs=[chatbot, source1, page1, source2, page2, source3, page3]
         )
+
+    # Launch the demo
     demo.queue().launch(debug=True)
 
 if __name__ == "__main__":