import os import PyPDF2 import gradio as gr # Import vectorstore and embeddings from langchain community package from langchain_community.vectorstores import FAISS from langchain_community.embeddings import HuggingFaceEmbeddings # Text splitter to break large documents into manageable chunks from langchain.text_splitter import CharacterTextSplitter # HF Inference client for running Mistral-7B chat completions from huggingface_hub import InferenceClient # ── Globals ─────────────────────────────────────────────────────────────────── index = None # FAISS index storing document embeddings retriever = None # Retriever to fetch relevant chunks current_pdf_name = None # Name of the currently loaded PDF pdf_text = None # Full text of the uploaded PDF # ── HF Inference client (token injected via Spaces secrets) ───────────────────── # Instantiate client for conversational endpoint (Mistral-7B-Instruct) client = InferenceClient(model="mistralai/Mistral-7B-Instruct-v0.3") # ── Embeddings ─────────────────────────────────────────────────────────────── # Use BGE embeddings from BAAI for vectorizing text chunks embeddings = HuggingFaceEmbeddings(model_name="BAAI/bge-base-en-v1.5") def process_pdf(pdf_file): """ 1. Reads and extracts text from each page of the uploaded PDF. 2. Splits the combined text into overlapping chunks for retrieval. 3. Builds a FAISS index over those chunks and initializes a retriever. Args: pdf_file: Filepath to the uploaded PDF. Returns: - PDF filename shown in UI - Status message with number of chunks - Enables the question input field """ global current_pdf_name, index, retriever, pdf_text # If no file uploaded, prompt the user if pdf_file is None: return None, "❌ Please upload a PDF file.", gr.update(interactive=False) # Save current filename for display and context current_pdf_name = os.path.basename(pdf_file.name) # Extract text from all pages with open(pdf_file.name, "rb") as f: reader = PyPDF2.PdfReader(f) pages = [page.extract_text() or "" for page in reader.pages] pdf_text = "\n\n".join(pages) # Combine page texts # Break text into 1,000-character chunks with 100-char overlap splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=100) chunks = splitter.split_text(pdf_text) # Build and store FAISS index for similarity search index = FAISS.from_texts(chunks, embeddings) # Create retriever configured to return top-2 most relevant chunks retriever = index.as_retriever(search_kwargs={"k": 2}) # Return filename, success status, and enable the question box status = f"✅ Indexed '{current_pdf_name}' — {len(chunks)} chunks" return current_pdf_name, status, gr.update(interactive=True) def ask_question(pdf_name, question): """ 1. Retrieves the top-k most relevant text chunks from the FAISS index. 2. Constructs a prompt combining those excerpts with the user question. 3. Calls the HF chat endpoint to generate an answer. Args: pdf_name: The displayed PDF filename (unused internally). question: The user's question about the document. Returns: The generated answer as a string. """ global retriever # Ensure a PDF is loaded first if index is None or retriever is None: return "❌ Please upload and index a PDF first." # Prompt user to type something if empty if not question.strip(): return "❌ Please enter a question." # Fetch relevant document chunks docs = retriever.get_relevant_documents(question) context = "\n\n".join(doc.page_content for doc in docs) # Prepare the conversational prompt prompt = ( "Use the following document excerpts to answer the question.\n\n" f"{context}\n\n" f"Question: {question}\n" "Answer:" ) # Run chat completion with the prompt as the user's message response = client.chat_completion( messages=[{"role": "user", "content": prompt}], max_tokens=128, temperature=0.5 ) # Parse assistant reply from the choices answer = response["choices"][0]["message"]["content"].strip() return answer def generate_summary(): """ Uses the first 2,000 characters of the loaded PDF text to ask the model for a concise summary. """ if not pdf_text: return "❌ Please upload and index a PDF first." # Shorten long docs to 2k chars for summarization prompt = ( "Please provide a concise summary of the following document:\n\n" f"{pdf_text[:2000]}..." ) response = client.chat_completion( messages=[{"role": "user", "content": prompt}], max_tokens=150, temperature=0.5 ) return response["choices"][0]["message"]["content"].strip() def extract_keywords(): """ Uses the first 2,000 characters to ask the model to extract key terms or concepts. """ if not pdf_text: return "❌ Please upload and index a PDF first." prompt = ( "Extract 10–15 key terms or concepts from the following document:\n\n" f"{pdf_text[:2000]}..." ) response = client.chat_completion( messages=[{"role": "user", "content": prompt}], max_tokens=60, temperature=0.5 ) return response["choices"][0]["message"]["content"].strip() def clear_interface(): """ Resets all global state back to None, and clears inputs in the UI. """ global index, retriever, current_pdf_name, pdf_text index = retriever = None current_pdf_name = pdf_text = None # Clear displayed filename and re-disable question input return None, "", gr.update(interactive=False) # ── Gradio UI ──────────────────────────────────────────────────────────────── theme = gr.themes.Soft(primary_hue="indigo", secondary_hue="blue") with gr.Blocks(theme=theme, css=""" .container { border-radius: 10px; padding: 15px; } .pdf-active { border-left: 3px solid #6366f1; padding-left: 10px; background-color: rgba(99,102,241,0.1); } .footer { text-align: center; margin-top: 30px; font-size: 0.8em; color: #666; } /* Center and enlarge the main heading */ .main-title { text-align: center; font-size: 64px; font-weight: bold; margin-bottom: 20px; } """) as demo: # Application title centered and bold gr.Markdown("
DocQueryAI
") with gr.Row(): with gr.Column(): gr.Markdown("## 📄 Document Input") # Display the name of the active PDF pdf_display = gr.Textbox(label="Active Document", interactive=False, elem_classes="pdf-active") # File upload widget for PDFs pdf_file = gr.File(file_types=[".pdf"], type="filepath") # Button to start processing upload_button = gr.Button("📤 Process Document", variant="primary") # Status text below the button status_box = gr.Textbox(label="Status", interactive=False) with gr.Column(): gr.Markdown("## ❓ Ask Questions") # Text area for user questions question_input = gr.Textbox(lines=3, placeholder="Enter your question here…") # Button to trigger Q&A ask_button = gr.Button("🔍 Ask Question", variant="primary") # Output textbox for the generated answer answer_output = gr.Textbox(label="Answer", lines=8, interactive=False) # Footer section with summary and keywords extraction with gr.Row(): summary_button = gr.Button("📋 Generate Summary", variant="secondary") summary_output = gr.Textbox(label="Summary", lines=4, interactive=False) keywords_button = gr.Button("🏷️ Extract Keywords", variant="secondary") keywords_output = gr.Textbox(label="Keywords", lines=4, interactive=False) # Clear everything clear_button = gr.Button("🗑️ Clear All", variant="secondary") gr.Markdown("
Powered by LangChain + Mistral 7B + FAISS | Gradio
") # Bind events to functions upload_button.click(process_pdf, [pdf_file], [pdf_display, status_box, question_input]) ask_button.click(ask_question, [pdf_display, question_input], answer_output) summary_button.click(generate_summary, [], summary_output) keywords_button.click(extract_keywords, [], keywords_output) clear_button.click(clear_interface, [], [pdf_file, pdf_display, question_input]) if __name__ == "__main__": # Launch the Gradio app, share=True exposes a public URL in Spaces demo.launch(debug=True, share=True)