"""STT provider: delegates to existing Memory Service /voice/stt.

Memory Service accepts: multipart/form-data audio file upload.
Returns: {text, model, language}

Fabric contract output: {text, segments[], language, meta}
"""
import base64
import logging
import os
from typing import Any, Dict

import httpx

logger = logging.getLogger("provider.stt_memory_service")

MEMORY_SERVICE_URL = os.getenv("MEMORY_SERVICE_URL", "http://memory-service:8000")
MAX_AUDIO_BYTES = int(os.getenv("STT_MAX_AUDIO_BYTES", str(25 * 1024 * 1024)))


async def _resolve_audio_bytes(payload: Dict[str, Any]) -> tuple[bytes, str, str, str]:
    """Return (raw_bytes, filename, source, content_type) from audio_b64 or audio_url."""
    audio_b64 = payload.get("audio_b64", "")
    audio_url = payload.get("audio_url", "")
    filename = payload.get("filename", "audio.wav")

    if audio_b64:
        raw = base64.b64decode(audio_b64)
        if len(raw) > MAX_AUDIO_BYTES:
            raise ValueError(f"Audio exceeds {MAX_AUDIO_BYTES} bytes")
        return raw, filename, "b64", "audio/wav"

    if audio_url:
        if audio_url.startswith(("file://", "/")):
            path = audio_url.replace("file://", "")
            with open(path, "rb") as f:
                raw = f.read()
            if len(raw) > MAX_AUDIO_BYTES:
                raise ValueError(f"Audio exceeds {MAX_AUDIO_BYTES} bytes")
            ext = path.rsplit(".", 1)[-1] if "." in path else "wav"
            return raw, f"audio.{ext}", "file", f"audio/{ext}"

        # HTTP URL — check Content-Length header first if available
        async with httpx.AsyncClient(timeout=30) as c:
            try:
                head_resp = await c.head(audio_url)
                content_length = int(head_resp.headers.get("content-length", 0))
                if content_length > MAX_AUDIO_BYTES:
                    raise ValueError(f"Audio URL Content-Length {content_length} exceeds {MAX_AUDIO_BYTES} bytes")
                content_type = head_resp.headers.get("content-type", "audio/wav")
            except httpx.HTTPError:
                content_type = "audio/wav"

            resp = await c.get(audio_url)
            resp.raise_for_status()
            raw = resp.content
            content_type = resp.headers.get("content-type", content_type)

        if len(raw) > MAX_AUDIO_BYTES:
            raise ValueError(f"Audio exceeds {MAX_AUDIO_BYTES} bytes")
        ext = content_type.split("/")[-1].split(";")[0] or "wav"
        return raw, f"audio.{ext}", "url", content_type

    raise ValueError("audio_b64 or audio_url is required")


async def transcribe(payload: Dict[str, Any]) -> Dict[str, Any]:
    """Fabric STT entry point — delegates to Memory Service.

    Payload:
        audio_url: str (http/file) — OR —
        audio_b64: str (base64)
        language: str (optional, e.g. "uk", "en")
        filename: str (optional, helps whisper detect format)

    Returns Fabric contract: {text, segments[], language, meta, provider, model}
    """
    language = payload.get("language")
    raw_bytes, filename, source, content_type = await _resolve_audio_bytes(payload)

    params = {}
    if language:
        params["language"] = language

    async with httpx.AsyncClient(timeout=90) as c:
        resp = await c.post(
            f"{MEMORY_SERVICE_URL}/voice/stt",
            files={"audio": (filename, raw_bytes, "audio/wav")},
            params=params,
        )
        resp.raise_for_status()
        data = resp.json()

    text = data.get("text", "")
    model_used = data.get("model", "faster-whisper")
    lang_detected = data.get("language", language or "")

    return {
        "text": text,
        "segments": [],
        "language": lang_detected,
        "meta": {
            "model": model_used,
            "provider": "memory_service",
            "engine": model_used,
            "service_url": MEMORY_SERVICE_URL,
            "source": source,
            "bytes": len(raw_bytes),
            "filename": filename,
            "content_type": content_type,
        },
        "provider": "memory_service",
        "model": model_used,
    }