aaa59558/bbb

1import gradio as gr2from transformers import AutoTokenizer, AutoModelForQuestionAnswering3from sentence_transformers import SentenceTransformer4import faiss5import os6import torch7 8print("🟢 App starting...")9 10# ✅ File to load11TXT_FILE = "149.txt"12if not os.path.exists(TXT_FILE):13    raise FileNotFoundError(f"❌ File not found: {TXT_FILE}")14 15# ✅ Chunking function16def chunk_text(text, max_words=300):17    words = text.split()18    return [" ".join(words[i:i + max_words]) for i in range(0, len(words), max_words)]19 20# ✅ Read and chunk the file21with open(TXT_FILE, "r", encoding="utf-8", errors="ignore") as f:22    text = f.read()23chunks = chunk_text(text)24 25# ✅ Embedding chunks26print("📐 Creating vector index...")27embedder = SentenceTransformer("all-MiniLM-L6-v2")28embeddings = embedder.encode(chunks)29index = faiss.IndexFlatL2(len(embeddings[0]))30index.add(embeddings)31 32# ✅ Load RoBERTa QA model33model_id = "deepset/roberta-base-squad2"34tokenizer = AutoTokenizer.from_pretrained(model_id)35model = AutoModelForQuestionAnswering.from_pretrained(model_id)36 37# ✅ Extractive QA function38def extract_answer(question, context):39    inputs = tokenizer(question, context, return_tensors="pt", truncation=True)40    with torch.no_grad():41        outputs = model(**inputs)42 43    start_idx = torch.argmax(outputs.start_logits)44    end_idx = torch.argmax(outputs.end_logits) + 145 46    if start_idx >= end_idx:47        return "❌ Couldn't extract a valid answer."48 49    answer = tokenizer.convert_tokens_to_string(50        tokenizer.convert_ids_to_tokens(inputs["input_ids"][0][start_idx:end_idx])51    )52    return answer53 54# ✅ Main chat function55def chat_with_bot(user_input):56    q_vec = embedder.encode([user_input])57    distances, indices = index.search(q_vec, k=3)58 59    if distances[0][0] > 1.5:60        return "❌ I couldn't find an answer in the document."61 62    # Try multiple chunks to find best answer63    best_answer = ""64    for i in indices[0]:65        answer = extract_answer(user_input, chunks[i])66        if answer and answer.strip() and "❌" not in answer:67            best_answer = answer68            break69 70    return f"🧠 {best_answer}" if best_answer else "❌ No confident answer found."71 72# ✅ Launch Gradio app73gr.Interface(74    fn=chat_with_bot,75    inputs=gr.Textbox(placeholder="Ask a factual question from 149.txt", lines=2),76    outputs="text",77    title="📚 Extractive QA Chatbot",78    description="Powered by RoBERTa-SQuAD2.0, answers directly from the document using span prediction."79).launch()80