config.py

"""Central configuration and performance knobs.

Everything tunable lives here and is overridable via environment variables, so you
can trade speed for accuracy without touching code (e.g. small Whisper model while
developing, large-v3-turbo for the demo).
"""

from __future__ import annotations

import os
from pathlib import Path

# --- Paths -------------------------------------------------------------------
ROOT = Path(__file__).resolve().parent
MODELS_DIR = ROOT / "models"
DATA_DIR = ROOT / "data"
DB_PATH = Path(os.getenv("ASA_DB_PATH", str(DATA_DIR / "jobs.db")))

DATA_DIR.mkdir(exist_ok=True)
MODELS_DIR.mkdir(exist_ok=True)


# --- Whisper (faster-whisper) ------------------------------------------------
# faster-whisper runs on CTranslate2: CPU or CUDA (no Apple Metal). On Mac, "int8"
# on CPU is the fast path. Use "small"/"base" while developing, "large-v3-turbo"
# for the demo — near-large accuracy at a fraction of the latency.
WHISPER_MODEL = os.getenv("ASA_WHISPER_MODEL", "medium")
WHISPER_DEVICE = os.getenv("ASA_WHISPER_DEVICE", "auto")          # auto|cpu|cuda
WHISPER_COMPUTE = os.getenv("ASA_WHISPER_COMPUTE", "int8")        # int8|int8_float16|float16
WHISPER_BEAM = int(os.getenv("ASA_WHISPER_BEAM", "1"))           # 1 = greedy, fastest
WHISPER_LANG = os.getenv("ASA_WHISPER_LANG", "en") or None        # None = autodetect


# --- Gemma 4 (transformers) — one multimodal model for STT + extraction ------
# E2B is the mobile-QAT checkpoint (2.3B effective params); bump to E4B for more
# accuracy headroom. Gated Google model: run `huggingface-cli login` and accept the
# license first.
GEMMA_MODEL_ID = os.getenv("ASA_GEMMA_MODEL", "google/gemma-4-E2B-it-qat-mobile-transformers")
GEMMA_DTYPE = os.getenv("ASA_GEMMA_DTYPE", "auto")       # auto|bfloat16|float16
GEMMA_DEVICE_MAP = os.getenv("ASA_GEMMA_DEVICE_MAP", "auto")
GEMMA_MAX_NEW_TOKENS = int(os.getenv("ASA_GEMMA_MAX_TOKENS", "1024"))
# Greedy by default: deterministic, best for form-filling + transcription. The card's
# recommended sampling (temp 1.0 / top_p 0.95 / top_k 64) only applies if you opt in.
GEMMA_SAMPLE = os.getenv("ASA_GEMMA_SAMPLE", "0") == "1"
GEMMA_TEMPERATURE = float(os.getenv("ASA_GEMMA_TEMPERATURE", "1.0"))
GEMMA_TOP_P = float(os.getenv("ASA_GEMMA_TOP_P", "0.95"))
GEMMA_TOP_K = int(os.getenv("ASA_GEMMA_TOP_K", "64"))
# Gemma's audio path is capped at 30s; longer notes fall back to faster-whisper.
GEMMA_AUDIO_MAX_SEC = float(os.getenv("ASA_GEMMA_AUDIO_MAX_SEC", "30"))


# --- llama.cpp (PRIMARY): Gemma 4 GGUF — text (GBNF grammar) + audio (mtmd) ---
# Preferred backend. Text extraction runs in-process via llama-cpp-python with a
# grammar so JSON is always valid. Audio runs through the llama-mtmd-cli binary
# (llama.cpp PR #21421); the mmproj MUST be BF16 per that PR. If the GGUF/mmproj/
# binary aren't present, the code falls back to the Gemma transformers path above.
LLAMA_MODEL_PATH = Path(
    os.getenv("ASA_LLAMA_MODEL", str(MODELS_DIR / "gemma-4-E2B-it-Q6_K.gguf"))
)
LLAMA_MMPROJ_PATH = Path(
    os.getenv("ASA_LLAMA_MMPROJ", str(MODELS_DIR / "mmproj-BF16.gguf"))
)
LLAMA_MTMD_CLI = os.getenv("ASA_LLAMA_MTMD_CLI", "llama-mtmd-cli")  # binary name/path
LLAMA_CTX = int(os.getenv("ASA_LLAMA_CTX", "4096"))
LLAMA_GPU_LAYERS = int(os.getenv("ASA_LLAMA_GPU_LAYERS", "-1"))     # -1 = all to GPU
LLAMA_THREADS = int(os.getenv("ASA_LLAMA_THREADS", str(os.cpu_count() or 4)))
LLAMA_TEMPERATURE = float(os.getenv("ASA_LLAMA_TEMPERATURE", "0.0"))
LLAMA_MAX_TOKENS = int(os.getenv("ASA_LLAMA_MAX_TOKENS", "1024"))


# --- Business defaults -------------------------------------------------------
CURRENCY = os.getenv("ASA_CURRENCY", "$")
BUSINESS_NAME = os.getenv("ASA_BUSINESS_NAME", "")  # shown on invoices, optional