web/config.py - main - Paff/declarant

web/config.py
62 строки3 KB
1pash1985-gif
feat: use Yandex Vision OCR as primary certificate scan fallback
09 авг 2026, 23:51
09 авг 2026, 23:517bf6a5c
"""Конфигурация веб-приложения."""
import os

# Пароль для входа (можно менять через переменную окружения)
ACCESS_PASSWORD = os.environ.get('DECLARANT_PASSWORD', 'declarant2026')

# Секрет для сессий Flask
SECRET_KEY = os.environ.get('SECRET_KEY', 'super-secret-key-change-me')

# File storage: use Docker volume (/app/data) in production, local dir in dev.
# REPORTS_DIR is set to /app/data/reports in Dockerfile; uploads go in /app/data/uploads.
_reports_dir = os.environ.get('REPORTS_DIR', '')
if _reports_dir:
    UPLOAD_FOLDER = os.path.join(os.path.dirname(_reports_dir), 'uploads')
else:
    UPLOAD_FOLDER = os.path.join(os.path.dirname(__file__), 'uploads')

# Максимальный размер загрузки (100 MB)
MAX_CONTENT_LENGTH = 100 * 1024 * 1024

# DeepSeek API
DEEPSEEK_API_KEY = os.environ.get('DEEPSEEK_API_KEY', 'sk-540f9e711cbb41e4b43241ab3d873ed5')
DEEPSEEK_BASE_URL = 'https://api.deepseek.com'
# Reasoning models (deepseek-v4-flash, deepseek-v4-pro) exhaust ALL max_tokens on
# reasoning_tokens for complex invoices, producing empty content. deepseek-chat is
# non-reasoning and produces reliable JSON output (~30s vs 100s+ for reasoning models).
DEEPSEEK_MODEL = os.environ.get('DEEPSEEK_MODEL', 'deepseek-chat')

# Kimi K3 API (Moonshot)
KIMI_API_KEY = os.environ.get('KIMI_API_KEY', 'sk-poRN6uORsFmQthQ76TsSgk2VLJbZbu6f0HoVldmRubGfFuLk')
KIMI_BASE_URL = 'https://api.moonshot.ai/v1'

# Разрешённые расширения
ALLOWED_EXTENSIONS = {'.pdf', '.xls', '.xlsx', '.xlsm', '.doc', '.zip', '.rar'}

# --- AI Ensemble: DeepSeek + Kimi ---
# Master switch. When False, the legacy single-model path is used.
ENSEMBLE_ENABLED = os.environ.get('ENSEMBLE_ENABLED', 'true').lower() in ('1', 'true', 'yes')

# Number of debate rounds when models disagree (0 = only parallel call + arbitration).
# Benchmark results (2026-07-27): rounds=1 is optimal (60-120s vs 250s+ for rounds=2).
ENSEMBLE_DEBATE_MAX_ROUNDS = int(os.environ.get('ENSEMBLE_DEBATE_MAX_ROUNDS', '1'))

# Per-task toggles. Only used when ENSEMBLE_ENABLED is True.
ENSEMBLE_ENABLE_PARSING = os.environ.get('ENSEMBLE_ENABLE_PARSING', 'true').lower() in ('1', 'true', 'yes')
ENSEMBLE_ENABLE_EUROPA = os.environ.get('ENSEMBLE_ENABLE_EUROPA', 'true').lower() in ('1', 'true', 'yes')
ENSEMBLE_ENABLE_IMPORT = os.environ.get('ENSEMBLE_ENABLE_IMPORT', 'true').lower() in ('1', 'true', 'yes')

# Total timeout (seconds) for the full ensemble call including debates.
# Benchmark results: 120s is sufficient for most files (avg 60-90s per file).
ENSEMBLE_TIMEOUT_SECONDS = int(os.environ.get('ENSEMBLE_TIMEOUT_SECONDS', '120'))

# AI ensemble as primary table-filling method (Таблица европа / Таблица импорт).
# When True, DeepSeek + Kimi jointly decide cell mapping before regex fallback.
ENSEMBLE_ENABLE_FILL = os.environ.get('ENSEMBLE_ENABLE_FILL', 'true').lower() in ('1', 'true', 'yes')

# Timeout for table-filling ensemble (larger — more items than single-invoice parse).
ENSEMBLE_FILL_TIMEOUT_SECONDS = int(os.environ.get('ENSEMBLE_FILL_TIMEOUT_SECONDS', '180'))

# Yandex Vision OCR (for scanned certificate pages fallback)
YANDEX_OCR_API_KEY = os.environ.get('YANDEX_OCR_API_KEY', '')
YANDEX_OCR_FOLDER_ID = os.environ.get('YANDEX_OCR_FOLDER_ID', '')