PTSRaiAgent/app/llm/stt.py

"""Speech-to-text for the patient's voice input (#116, slide 11: record → check the text → send).

The service only transcribes: the text goes back to the patient to check and edit, and reaches the dialog as an
ordinary text message — safety, `scrub()` and the LLM gateway apply there. Audio is never stored or logged.

* `OpenRouterSTT` — an audio-capable model via OpenRouter `/chat/completions` (`input_audio`). TEMPORARY, synthetic
  data only (ADR-014): voice is personal data that `scrub()` cannot depersonalise, and OpenRouter is a foreign
  service (152-FZ). Real patients need an RF provider (SaluteSpeech, #121) behind the same `SpeechToText` contract.
* `MockSTT` — offline, returns a fixed synthetic phrase (ADR-004: the demo works without network).
"""

import asyncio
import base64
import logging
import struct
from functools import lru_cache
from pathlib import Path
from typing import Protocol

import httpx
import yaml
from pydantic import BaseModel

from app.config import Settings

log = logging.getLogger(__name__)

_RETRY_STATUSES = {429, 500, 502, 503, 504}
_ATTEMPTS = 3


class SttTexts(BaseModel):
    system: str
    user: str
    silence_marker: str
    mock_transcript: str


@lru_cache
def load_texts(path: Path | None = None) -> SttTexts:
    from app.config import get_settings

    path = path or get_settings().config_dir / "texts" / "stt.yaml"
    return SttTexts.model_validate(yaml.safe_load(path.read_text(encoding="utf-8")))


class SttError(RuntimeError):
    """The provider failed (network, HTTP error, bad response)."""


class SttNoSpeech(SttError):
    """The recording holds no recognisable speech."""


class InvalidAudio(ValueError):
    pass


def wav_seconds(audio: bytes) -> float:
    """Duration of a PCM WAV file; raises `InvalidAudio` for anything else."""
    if len(audio) < 44 or audio[:4] != b"RIFF" or audio[8:12] != b"WAVE":
        raise InvalidAudio("not a WAV file")
    pos, byte_rate = 12, 0
    while pos + 8 <= len(audio):
        chunk_id, size = audio[pos : pos + 4], struct.unpack_from("<I", audio, pos + 4)[0]
        if chunk_id == b"fmt " and size >= 16:
            byte_rate = struct.unpack_from("<I", audio, pos + 16)[0]
        elif chunk_id == b"data":
            if not byte_rate:
                break
            return min(size, len(audio) - pos - 8) / byte_rate
        pos += 8 + size + (size & 1)
    raise InvalidAudio("malformed WAV")


class SpeechToText(Protocol):
    name: str

    async def transcribe(self, audio: bytes, fmt: str = "wav") -> str: ...


class MockSTT:
    name = "mock"

    async def transcribe(self, audio: bytes, fmt: str = "wav") -> str:
        return load_texts().mock_transcript


class OpenRouterSTT:
    name = "openrouter"

    def __init__(self, settings: Settings, client: httpx.AsyncClient | None = None) -> None:
        self._key = settings.openrouter_api_key
        self._model = settings.stt_model
        self._url = settings.openrouter_base_url.rstrip("/") + "/chat/completions"
        self._timeout = settings.stt_timeout_s
        self._proxy = settings.openrouter_proxy_url or None
        self._client = client

    async def transcribe(self, audio: bytes, fmt: str = "wav") -> str:
        texts = load_texts()
        content = [
            {"type": "text", "text": texts.user},
            {"type": "input_audio", "input_audio": {"data": base64.b64encode(audio).decode(), "format": fmt}},
        ]
        body = {
            "model": self._model,
            "messages": [{"role": "system", "content": texts.system}, {"role": "user", "content": content}],
            "temperature": 0,
            "max_tokens": 800,
        }
        data = await self._post(body)
        try:
            text = (data["choices"][0]["message"]["content"] or "").strip()
        except (KeyError, IndexError, TypeError) as exc:
            raise SttError("bad response") from exc
        text = text.strip("«»\"' \n")
        if not text or text.lower() == texts.silence_marker.lower():
            raise SttNoSpeech("no speech")
        return text

    async def _post(self, body: dict) -> dict:
        headers = {"Authorization": f"Bearer {self._key}", "X-Title": "PTSRaiAgent prototype (synthetic data)"}
        for attempt in range(_ATTEMPTS):
            try:
                if self._client is not None:
                    resp = await self._client.post(self._url, json=body, headers=headers, timeout=self._timeout)
                else:
                    async with httpx.AsyncClient(timeout=self._timeout, proxy=self._proxy) as client:
                        resp = await client.post(self._url, json=body, headers=headers)
            except httpx.HTTPError as exc:
                if attempt + 1 == _ATTEMPTS:
                    raise SttError(type(exc).__name__) from exc
                await asyncio.sleep(0.5 * 2**attempt)
                continue
            if resp.status_code in _RETRY_STATUSES and attempt + 1 < _ATTEMPTS:
                log.warning("stt openrouter http %s, retry %d", resp.status_code, attempt + 1)
                await asyncio.sleep(0.5 * 2**attempt)
                continue
            if resp.status_code != 200:
                # Provider error bodies carry no patient data — log the message to make failures diagnosable.
                log.error("stt openrouter http %s: %s", resp.status_code, resp.text[:200])
                raise SttError(f"http {resp.status_code}")
            try:
                return resp.json()
            except ValueError as exc:
                raise SttError("bad json") from exc
        raise SttError("retries exhausted")


def get_stt(settings: Settings) -> SpeechToText | None:
    """The configured provider; `None` = voice input is off. `auto` = OpenRouter if a key is set, else mock."""
    mode = settings.stt_provider
    if mode == "auto":
        mode = "openrouter" if settings.openrouter_api_key else "mock"
    if mode == "openrouter":
        return OpenRouterSTT(settings) if settings.openrouter_api_key else None
    if mode == "mock":
        return MockSTT()
    return None