+ Strea,

app.py

@@ -4,7 +4,7 @@ import asyncio
 from itertools import cycle
 from core.utils import generate_first_question
 from core.mbti_analyzer import analyze_mbti
-from core.interviewer import generate_question
+from core.interviewer import stream_question  # ✅ теперь используем потоковую версию
 
 # --------------------------------------------------------------
 # 🌀 Асинхронная анимация "Thinking..."
@@ -15,11 +15,12 @@ async def async_loader(update_fn, delay=0.15):
         update_fn(f"💭 Interviewer is thinking... {frame}")
         await asyncio.sleep(delay)
 
-
 # --------------------------------------------------------------
 # ⚙️ Основная логика
 # --------------------------------------------------------------
 def analyze_and_ask(user_text, prev_count):
+    """Основная функция — анализирует ответ и генерирует следующий вопрос (потоково)."""
+
     if not user_text.strip():
         yield "⚠️ Please enter your answer.", "", prev_count
         return
@@ -33,21 +34,21 @@ def analyze_and_ask(user_text, prev_count):
     # мгновенный отклик
     yield "⏳ Analyzing personality...", "💭 Interviewer is thinking... ⠋", counter
 
-    # анализ MBTI
+    # анализ MBTI (также потоковый)
     mbti_gen = analyze_mbti(user_text)
     mbti_text = ""
     for chunk in mbti_gen:
         mbti_text = chunk
         yield mbti_text, "💭 Interviewer is thinking... ⠙", counter
 
-    # генерация вопроса новой моделью (без инструкций)
+    # генерация вопроса новой моделью (потоково)
     try:
-        question = generate_question()
+        partial_question = ""
+        for piece in stream_question():  # 👈 здесь идёт токен-за-токен поток
+            partial_question = piece
+            yield mbti_text, partial_question, counter
     except Exception as e:
-        question = f"⚠️ Question generator error: {e}"
-
-    yield mbti_text, question, counter
-
+        yield mbti_text, f"⚠️ Question generator error: {e}", counter
 
 # --------------------------------------------------------------
 # 🧱 Интерфейс Gradio
@@ -55,7 +56,8 @@ def analyze_and_ask(user_text, prev_count):
 with gr.Blocks(theme=gr.themes.Soft(), title="MBTI Personality Interviewer") as demo:
     gr.Markdown(
         "## 🧠 MBTI Personality Interviewer\n"
-        "Определи личностный тип и получи случайные вопросы MBTI категории."
+        "Определи личностный тип и получи вопросы из разных категорий MBTI.\n\n"
+        "_Теперь с потоковой генерацией вопросов._"
     )
 
     with gr.Row():
@@ -63,25 +65,25 @@ with gr.Blocks(theme=gr.themes.Soft(), title="MBTI Personality Interviewer") as
             inp = gr.Textbox(
                 label="Ваш ответ",
                 placeholder="Например: I enjoy working with people and organizing events.",
-                lines=4
+                lines=4,
             )
             btn = gr.Button("Анализировать и задать новый вопрос", variant="primary")
         with gr.Column(scale=1):
             mbti_out = gr.Textbox(label="📊 Анализ MBTI", lines=4)
-            interviewer_out = gr.Textbox(label="💬 Следующий вопрос", lines=3)
+            interviewer_out = gr.Textbox(label="💬 Следующий вопрос (streaming)", lines=3)
             progress = gr.Textbox(label="⏳ Прогресс", value="0/8")
 
     btn.click(
         analyze_and_ask,
         inputs=[inp, progress],
         outputs=[mbti_out, interviewer_out, progress],
-        show_progress=True
+        show_progress=True,
     )
 
     demo.load(
         lambda: ("", generate_first_question(), "0/8"),
         inputs=None,
-        outputs=[mbti_out, interviewer_out, progress]
+        outputs=[mbti_out, interviewer_out, progress],
     )
 
 demo.queue(max_size=32).launch(server_name="0.0.0.0", server_port=7860)

core/interviewer.py

@@ -1,94 +1,44 @@
 # core/interviewer.py
-"""
-🇬🇧 Interviewer logic module (no instructions)
-Generates random MBTI-style questions using the fine-tuned model.
-
-🇷🇺 Модуль интервьюера.
-Использует fine-tuned модель для генерации вопросов без инструкций.
-"""
-
-import random
-import re
 import torch
-from transformers import AutoModelForSeq2SeqLM, T5Tokenizer
+import threading
+from transformers import AutoModelForSeq2SeqLM, T5Tokenizer, TextIteratorStreamer
 
-# --------------------------------------------------------------
-# 1️⃣ Настройки модели
-# --------------------------------------------------------------
 QG_MODEL = "f3nsmart/ft-flan-t5-base-qgen_v2"
 
-
-# ✅ Принудительно используем оригинальный SentencePiece-токенайзер
 tokenizer = T5Tokenizer.from_pretrained(QG_MODEL, use_fast=False)
 model = AutoModelForSeq2SeqLM.from_pretrained(QG_MODEL)
-
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 model.to(device).eval()
-print(f"✅ Loaded interviewer model (slow tokenizer): {QG_MODEL}")
-print(f"Device set to use {device}")
-
-# --------------------------------------------------------------
-# 2️⃣ Seed-промпты (без инструкций)
-# --------------------------------------------------------------
-
-# --------------------------------------------------------------
-# 2️⃣ Тематические seed-промпты (по осям MBTI, но без прямого упоминания MBTI)
-# --------------------------------------------------------------
 
-BASE_INSTRUCTION = (
-    "Generate one natural, open-ended question about human thinking, emotions, or decision-making. "
-    "Avoid mentioning MBTI or personality types directly. "
-    "Do not ask what type the person belongs to. "
-    "You may include ideas related to intuition, logic, feelings, perception, or judgment naturally."
-)
+print(f"✅ Loaded interviewer model (streaming ready): {QG_MODEL}")
 
-PROMPTS = [
-    f"{BASE_INSTRUCTION} Explore how people usually recharge their energy and interact with others.",
-    f"{BASE_INSTRUCTION} Explore the difference between noticing small details and seeing the bigger picture.",
-    f"{BASE_INSTRUCTION} Ask about trusting intuition versus relying on concrete evidence in daily life.",
-    f"{BASE_INSTRUCTION} Ask about what typically inspires or motivates someone to take action.",
-    f"{BASE_INSTRUCTION} Create a question about balancing emotions and logic when making decisions.",
-    f"{BASE_INSTRUCTION} Write about preferences between careful planning and spontaneous action.",
-    f"{BASE_INSTRUCTION} Explore how individuals deal with uncertainty or unexpected changes.",
-    f"{BASE_INSTRUCTION} Ask about understanding other people’s emotions or empathy in relationships.",
-    f"{BASE_INSTRUCTION} Create a question about staying organized versus adapting flexibly to new situations.",
-    f"{BASE_INSTRUCTION} Explore curiosity, creativity, and how people find meaning in what they do."
-]
-
-
-# --------------------------------------------------------------
-# 3️⃣ Очистка текста
-# --------------------------------------------------------------
-def _clean_question(text: str) -> str:
-    """Берёт первую фразу с '?'"""
-    text = text.strip()
-    m = re.search(r"(.+?\?)", text)
-    if m:
-        text = m.group(1)
-    text = text.replace("\n", " ").strip()
-    if len(text.split()) < 3:
-        text = text.capitalize()
-        if not text.endswith("?"):
-            text += "?"
-    return text
-
-
-# --------------------------------------------------------------
-# 4️⃣ Генерация вопроса
-# --------------------------------------------------------------
-def generate_question(user_id: str = "default_user", **kwargs) -> str:
-    """Генерирует один MBTI-вопрос без инструкций"""
-    prompt = random.choice(PROMPTS)
+# обычная версия (если нужно fallback)
+def generate_question(prompt: str = "Generate one thoughtful question.") -> str:
     inputs = tokenizer(prompt, return_tensors="pt", truncation=True).to(device)
     with torch.no_grad():
-        out = model.generate(
-            **inputs,
-            do_sample=True,
-            top_p=0.9,
-            temperature=1.1,
-            top_k = 60,
-            repetition_penalty=1.5,
-            max_new_tokens=80,
-        )
-    text = tokenizer.decode(out[0], skip_special_tokens=True)
-    return _clean_question(text)
+        output = model.generate(**inputs, max_new_tokens=80)
+    return tokenizer.decode(output[0], skip_special_tokens=True)
+
+# потоковая версия
+def stream_question(prompt: str = "Generate one thoughtful question."):
+    inputs = tokenizer(prompt, return_tensors="pt", truncation=True).to(device)
+    streamer = TextIteratorStreamer(tokenizer, skip_special_tokens=True)
+    generation_kwargs = dict(
+        **inputs,
+        streamer=streamer,
+        max_new_tokens=80,
+        do_sample=True,
+        top_p=0.9,
+        temperature=1.1,
+        top_k=60,
+        repetition_penalty=1.3,
+    )
+
+    # модель работает в отдельном потоке
+    thread = threading.Thread(target=model.generate, kwargs=generation_kwargs)
+    thread.start()
+
+    partial = ""
+    for new_text in streamer:
+        partial += new_text
+        yield partial