Back to normal app.py v1.1

app.py

@@ -1,58 +1,56 @@
-import asyncio
+# app.py
 import gradio as gr
 from core.utils import generate_first_question
 from core.mbti_analyzer import analyze_mbti
-from core.interviewer import generate_next_question
+from core.interviewer import generate_question
 
-
-# ---- Адаптер ----
-def analyze_and_ask_sync(user_text, prev_count, user_id="default_user"):
-    """Синхронный адаптер для Gradio"""
-    return asyncio.run(analyze_and_ask(user_text, prev_count, user_id))
-
-
-# ---- Асинхронная логика ----
-async def analyze_and_ask(user_text, prev_count, user_id="default_user"):
+def analyze_and_ask(user_text, prev_count):
+    """Пошаговый генератор — стриминг без async и без streaming=True."""
     if not user_text.strip():
-        return "⚠️ Введите ответ.", "", prev_count
+        yield "⚠️ Please enter your answer.", "", prev_count
+        return
 
     try:
         n = int(prev_count.split("/")[0]) + 1
     except Exception:
         n = 1
-    counter = f"{n}/16"
+    counter = f"{n}/30"
 
+    # 1️⃣ Шаг 1 — анализ
+    mbti_gen = analyze_mbti(user_text)
     mbti_text = ""
-    for part in analyze_mbti(user_text):
-        mbti_text = part
-
-    next_q_data = await generate_next_question(user_id, user_text)
-    if next_q_data["completed"]:
-        next_question = "✅ Interview finished! All 16 personality categories covered."
-    else:
-        next_question = f"({next_q_data['category']}) {next_q_data['question']}"
-
-    return mbti_text, next_question, counter
-
-
-# ---- Gradio UI ----
-with gr.Blocks(theme=gr.themes.Soft(), title="MBTI Interviewer AI") as demo:
-    gr.Markdown("## 🧠 MBTI Personality Interviewer\nАнализ и вопросы по 16 категориям MBTI.")
+    for chunk in mbti_gen:
+        mbti_text = chunk
+        yield mbti_text, "💭 Interviewer is thinking...", counter
+
+    # 2️⃣ Шаг 2 — вопрос
+    interviewer_gen = generate_question("default_user", user_text)
+    next_q = ""
+    for chunk in interviewer_gen:
+        next_q = chunk
+        yield mbti_text, next_q, counter
+
+# --------------------------------------------------------------
+# Gradio интерфейс
+# --------------------------------------------------------------
+with gr.Blocks(theme=gr.themes.Soft(), title="MBTI Personality Interviewer") as demo:
+    gr.Markdown("## 🧠 MBTI Personality Interviewer\nОпредели личностный тип и получи следующий вопрос от интервьюера.")
 
     with gr.Row():
         with gr.Column(scale=1):
-            inp = gr.Textbox(label="Ваш ответ", placeholder="Например: I enjoy organizing group projects.", lines=4)
-            btn = gr.Button("Отправить", variant="primary")
+            inp = gr.Textbox(
+                label="Ваш ответ",
+                placeholder="Например: I enjoy working with people and organizing events.",
+                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)
-            progress = gr.Textbox(label="⏳ Прогресс", value="0/16")
+            interviewer_out = gr.Textbox(label="💬 Следующий вопрос от интервьюера", lines=3)
+            progress = gr.Textbox(label="⏳ Прогресс", value="0/30")
 
-    # Здесь вызываем адаптер
-    btn.click(analyze_and_ask_sync, inputs=[inp, progress], outputs=[mbti_out, interviewer_out, progress])
+    btn.click(analyze_and_ask, inputs=[inp, progress], outputs=[mbti_out, interviewer_out, progress])
 
-    demo.load(lambda: ("", generate_first_question(), "0/16"),
-              inputs=None, outputs=[mbti_out, interviewer_out, progress])
+    demo.load(lambda: ("", generate_first_question(), "0/30"), inputs=None, outputs=[mbti_out, interviewer_out, progress])
 
-if __name__ == "__main__":
-    demo.launch()
+demo.queue(max_size=20).launch(server_name="0.0.0.0", server_port=7860)

core/interviewer.py

@@ -1,102 +1,55 @@
 # core/interviewer.py
-import asyncio, random, uuid
+import asyncio
 from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
 
-# Используем Phi-3-mini-instruct
-INTERVIEWER_MODEL = "microsoft/Phi-3-mini-4k-instruct"
+INTERVIEWER_MODEL = "f3nsmart/TinyLlama-MBTI-Interviewer-LoRA"
 
 tokenizer = AutoTokenizer.from_pretrained(INTERVIEWER_MODEL)
-model = AutoModelForCausalLM.from_pretrained(
-    INTERVIEWER_MODEL,
-    torch_dtype="auto",
-    device_map="auto"
-)
+model = AutoModelForCausalLM.from_pretrained(INTERVIEWER_MODEL, torch_dtype="auto", device_map="auto")
 
 llm_pipe = pipeline(
     "text-generation",
     model=model,
     tokenizer=tokenizer,
-    max_new_tokens=80,
+    max_new_tokens=70,
     temperature=0.7,
     top_p=0.9,
 )
 
-# Память о пользователе
-user_sessions = {}
-
-# 16 категорий MBTI (можно адаптировать под твои .json)
-MBTI_CATEGORIES = [
-    "ENFJ", "ENFP", "ENTJ", "ENTP",
-    "ESFJ", "ESFP", "ESTJ", "ESTP",
-    "INFJ", "INFP", "INTJ", "INTP",
-    "ISFJ", "ISFP", "ISTJ", "ISTP"
-]
+user_memory = {}
 
 def clean_question(text: str) -> str:
-    """Очищает вопрос от инструкций."""
-    text = text.strip().split("\n")[0]
-    text = text.strip('"').strip("'")
-    for bad in ["user:", "assistant:", "system:", "instruction"]:
+    text = text.strip().split("\n")[0].strip('"').strip("'")
+    bad_tokens = ["user:", "assistant:", "instruction", "interviewer", "system:"]
+    for bad in bad_tokens:
         if bad.lower() in text.lower():
             text = text.split(bad)[-1].strip()
     if not text.endswith("?"):
         text += "?"
-    return text
-
-async def generate_next_question(user_id: str, user_text: str = "") -> dict:
-    """
-    Генерирует следующий вопрос по категории.
-    """
-    session = user_sessions.get(user_id, {
-        "history": [],
-        "category_index": 0,
-        "completed": False
-    })
-
-    # Проверяем, не закончились ли категории
-    if session["category_index"] >= len(MBTI_CATEGORIES):
-        session["completed"] = True
-        user_sessions[user_id] = session
-        return {
-            "question": None,
-            "category": None,
-            "completed": True
-        }
+    return text if len(text.split()) > 3 else "What do you usually enjoy doing in your free time?"
 
-    current_category = MBTI_CATEGORIES[session["category_index"]]
-
-    # Составляем контекст (историю последних 2 ответов)
-    history = session["history"][-4:]
-    formatted_history = "\n".join(
-        [f"Q{i//2+1}: {history[i]}\nA{i//2+1}: {history[i+1]}" for i in range(0, len(history), 2)]
-    )
+def generate_question(user_id: str, user_text: str):
+    """Генератор вопроса (стриминг)."""
+    prev_qs = user_memory.get(user_id, [])
+    prev_joined = "; ".join(prev_qs) if prev_qs else "None"
 
     prompt = (
-        f"You are an MBTI interviewer.\n"
-        f"Current personality category: {current_category}\n\n"
-        f"Previous dialogue:\n{formatted_history}\n\n"
-        f"Now generate one new, open-ended question (starting with What, Why, How, or When) "
-        f"to assess this category.\n"
-        f"Do NOT repeat or rephrase any previous question.\n"
-        f"Return only the question text."
+        f"The following is an MBTI personality interview.\n"
+        f"User: {user_text}\n"
+        f"Interviewer: ask one new, open-ended question starting with 'What', 'Why', 'How', or 'When'. "
+        f"Avoid repeating or rephrasing previous questions.\n"
+        f"Previous questions: {prev_joined}\n"
+        f"Interviewer:"
     )
 
-    # Генерация вопроса
-    loop = asyncio.get_event_loop()
-    raw = await loop.run_in_executor(None, lambda: llm_pipe(prompt)[0]["generated_text"])
+    yield "💭 Interviewer is thinking..."
+    raw = llm_pipe(prompt)[0]["generated_text"]
     question = clean_question(raw)
 
-    # Уникальный ID вопроса
-    q_id = str(uuid.uuid4())[:8]
-
-    # Обновляем состояние пользователя
-    session["history"].extend([question, user_text])
-    session["category_index"] += 1
-    user_sessions[user_id] = session
+    valid_starts = ("What", "Why", "How", "When")
+    if not question.startswith(valid_starts):
+        question = "What motivates you to do the things you enjoy most?"
 
-    return {
-        "id": q_id,
-        "question": question,
-        "category": current_category,
-        "completed": False
-    }
+    prev_qs.append(question)
+    user_memory[user_id] = prev_qs[-10:]
+    yield question