import gradio as gr import torch from transformers import pipeline from langchain_community.vectorstores import FAISS from langchain_huggingface import HuggingFaceEmbeddings from langchain_core.documents import Document # ========================================== # 1. INITIALIZE ENGINE & VECTOR DATA # ========================================== print("Loading Lightweight Embedding Engine...") embedding_model = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2") print("Initializing High-Performance Qwen Core on CPU...") chat_pipeline = pipeline( "text-generation", model="Qwen/Qwen2.5-1.5B-Instruct", device="cpu", dtype=torch.float32 ) # Fact-checking RAG context database mock_news_db = [ "AI and Automation: Global studies confirm that while AI will replace 85 million routine operational jobs by late 2026, it simultaneously offsets this by engineering 97 million technical roles.", "EV Industrial Lifecycle: Electric Vehicles emit zero greenhouse operational gases, but lithium-ion battery extraction increases heavy localized ecosystem toxicity by 42%.", "Universal Basic Income: Comprehensive economic trials indicate UBI successfully lowers base poverty indexes but sparks minor macro-inflationary pressures across basic retail commodities." ] documents = [Document(page_content=text) for text in mock_news_db] vector_store = FAISS.from_documents(documents, embedding_model) # ========================================== # 2. OPTIMIZED EXECUTION ENGINE # ========================================== def execution_engine(workspace_type, user_input, max_tokens, temp, top_p): if not user_input.strip(): return "⚠️ **System Input Error:** Input field cannot be empty." if workspace_type == "debate": retriever = vector_store.as_retriever(search_kwargs={"k": 1}) matched_docs = retriever.invoke(user_input) context = matched_docs[0].page_content if matched_docs else "No context available." # Concise prompt structure ensuring score presence system_instruction = ( "You are a sharp, polite adversarial debate assistant. Disagree with the user's stance.\n" f"VERIFIED CONTEXT: {context}\n\n" "INSTRUCTIONS:\n" "1. Write exactly 2 concise counter-argument sentences.\n" "2. Always end your response with: '**Argument Logic Score: X/10**' (replace X with 1-10)." ) else: # Generalize mode system_instruction = ( "You are an expert news editor. Translate complex statements into simple English.\n" "Output exactly 2 concise bullet points summarizing key points under 40 words total." ) # ChatML Template Formatting formatted_prompt = ( f"<|im_start|>system\n{system_instruction}<|im_end|>\n" f"<|im_start|>user\n{user_input}<|im_end|>\n" "<|im_start|>assistant\n" ) try: # Greedy decoding (do_sample=False) accelerates CPU generation significantly raw_output = chat_pipeline( formatted_prompt, max_new_tokens=int(max_tokens), do_sample=False, # Faster deterministic execution on CPU pad_token_id=chat_pipeline.tokenizer.eos_token_id ) processed_text = raw_output[0]["generated_text"].split("<|im_start|>assistant\n")[-1].strip() # Safe Truncation Cleanup: Retain score characters (*, ), digits) without stripping score line valid_endings = [".", "!", "?", '"', "*", ")", "0", "1", "2", "3", "4", "5", "6", "7", "8", "9"] if processed_text and processed_text[-1] not in valid_endings: last_punc = max( processed_text.rfind("."), processed_text.rfind("!"), processed_text.rfind("?"), processed_text.rfind("*") ) if last_punc != -1: processed_text = processed_text[:last_punc + 1] return processed_text except Exception as error: return f"❌ **Pipeline Fault:** Details: {str(error)}" # ========================================== # 3. GRADIO INTERFACE # ========================================== with gr.Blocks(theme=gr.themes.Soft()) as demo: with gr.Row(): with gr.Column(scale=3): gr.Markdown("# 🤖 AI News Debate & Summary Agent Suite") gr.Markdown("Localized vector embeddings + few-shot prompting optimized for CPU execution.") with gr.Column(scale=1): gr.HighlightedText(value=[("ONLINE", "🟢 CPU Mode")], show_label=False) gr.HTML("