# Установка необходимых библиотек
!pip -q install sentence-transformers gradio scikit-learn

import os, json, math, random, re, pathlib, textwrap
from typing import List, Dict, Tuple, Optional
import numpy as np
import requests

GOOGLE_DRIVE_JSON_URL = "https://drive.google.com/uc?export=download&id=..."

response = requests.get(GOOGLE_DRIVE_JSON_URL)
response.raise_for_status()
data = response.json()
# дальше FileIndex разбирает data и вытаскивает текст

class LLMClient:
    def __init__(self):
        self.yc_key = os.getenv("YC_API_KEY", "").strip()
        self.yc_folder = os.getenv("YC_FOLDER_ID", "").strip()

    def chat_messages(self, messages, model_yandex="yandexgpt-lite",
                      temperature=0.3, max_tokens=1500) -> str:
        try:
            if self.yc_key and self.yc_folder:
                return self._yandex_chat(messages, model_yandex, temperature, max_tokens)
        except Exception:
            return self._fallback_answer(messages)
        return self._fallback_answer(messages)

def _encode(texts: List[str]) -> np.ndarray:
    if _USE_ST:
        v = _MODEL.encode(texts, normalize_embeddings=True, show_progress_bar=False)
        return np.array(v, dtype=np.float32)
    # запасной режим TF-IDF ...

def _cosine(a: np.ndarray, b: np.ndarray) -> float:
    na = np.linalg.norm(a) + 1e-9
    nb = np.linalg.norm(b) + 1e-9
    return float(np.dot(a, b) / (na * nb))

class FileIndex:
    def search(self, query: str, top_k: int = 3) -> List[Dict]:
        if not self.docs:
            return []
        qv = _encode([query])[0]
        scores = [_cosine(qv, self.vectors[i]) for i in range(len(self.docs))]
        idx = np.argsort(scores)[::-1][:top_k]
        return [{"score": float(scores[i]), **self.docs[i]} for i in idx]

def pal_run(code: str) -> str:
    loc = {}
    try:
        exec(code, {"__builtins__": __builtins__, "math": math, "np": np}, loc)
        return str(loc.get("result", ""))  # ожидается, что код положит что-то в result
    except Exception as e:
        return f"(Ошибка PAL: {e})"

EXPERT_PROFILES = {
    "Экономист": "...",
    "Переводчик": "...",
    "Международник": "...",
    "Специалист по работе с данными": "..."
}

BASE_SYSTEM = (
    "Отвечайте лаконично. Используйте ФАКТЫ при их наличии..."
)

class NanoAssistant:
    def __init__(self, expert_key: str = "Экономист"):
        self.llm = LLMClient()
        self.index = FileIndex()
        self.index.build()
        self.expert_key = expert_key

    def set_expert(self, expert_key: str):
        self.expert_key = expert_key

    def answer(self, query: str, try_pal: bool = True, use_fewshot: bool = True) -> str:
        facts = self._build_facts_block(query, top_k=3)
        pal_out = ""
        # логика PAL и сборки messages
        return self.llm.chat_messages(messages)

import gradio as gr

assistant = NanoAssistant(expert_key="Экономист")

def ui_answer(query, try_pal, expert_key, use_fewshot):
    assistant.set_expert(expert_key)
    return assistant.answer(query, try_pal=try_pal, use_fewshot=use_fewshot)

with gr.Blocks(title="AI Nano ассистент") as demo:
    gr.Markdown("## AI Nano\nВыберите профиль эксперта и задайте вопрос.")
    with gr.Row():
        expert_key = gr.Dropdown(choices=list(EXPERT_PROFILES.keys()),
                                 value="Экономист",
                                 label="Профиль эксперта")
        use_fewshot = gr.Checkbox(label="Включить примеры", value=True)
        try_pal = gr.Checkbox(label="Включить PAL (расчеты на Python)", value=True)
    inp = gr.Textbox(label="Ваш вопрос", lines=6)
    out = gr.Markdown()
    gr.Button("Ответить").click(ui_answer,
                                inputs=[inp, try_pal, expert_key, use_fewshot],
                                outputs=[out])

demo.launch(share=False)