get_wechat/chatlog_fastAPI/services/media_parser.py

import asyncio
import base64
import logging

import httpx
from fastapi import HTTPException

from services.ai_client import get_client_for
from services.media_resolver import resolve_media
from services.runtime_settings import get_ai_settings

log = logging.getLogger(__name__)

# 语音异步 ASR 默认网关（阿里云）。voice_base_url 为空时回退到此；
# 提交任务/轮询的子路径由代码自动拼接，配置只需填到 .../api/v1 这一层。
DEFAULT_ASR_BASE_URL = "https://dashscope.aliyuncs.com/api/v1"


async def parse_media(kind: str, key: str) -> dict:
    """
    Parse one chatlog media object into text.

    kind: voice, image, or video.
    key: chatlog media key.
    """
    if kind not in {"voice", "image", "video"}:
        raise HTTPException(400, "不支持的媒体类型")
    if not key:
        raise HTTPException(400, "媒体 key 不能为空")

    ai = await get_ai_settings()
    # voice/vision 各自有独立 url/key，为空则回退全局 ai_api_key
    if kind == "voice":
        if not (ai.get("voice_api_key") or ai.get("ai_api_key")):
            raise HTTPException(503, "AI 服务未配置，请在设置页填写语音密钥或 AI API Key")
        if not ai.get("voice_model"):
            raise HTTPException(503, "语音模型未配置，请在设置页填写语音模型名称，例如 paraformer-v2")
    if kind in ("image", "video"):
        if not (ai.get("vision_api_key") or ai.get("ai_api_key")):
            raise HTTPException(503, "AI 服务未配置，请在设置页填写视觉密钥或 AI API Key")
        if not ai.get("vision_model"):
            raise HTTPException(503, "视觉模型未配置，请在设置页填写视觉模型名称，例如 qwen-vl-plus")

    media = await resolve_media(kind, key)
    if kind == "voice":
        return {"text": await _parse_voice(media.bytes, media.content_type)}
    return {"text": await _parse_visual(kind, media.bytes, media.content_type)}


def _audio_mime(content_type: str) -> str:
    """由 chatlog 返回的 content_type 推断音频 MIME（用于 data URI）。"""
    ct = content_type.lower()
    if "silk" in ct or "x-silk" in ct:
        return "audio/silk"
    if "amr" in ct:
        return "audio/amr"
    if "ogg" in ct or "opus" in ct:
        return "audio/ogg"
    if "wav" in ct:
        return "audio/wav"
    return "audio/mpeg"


def _asr_json(resp: httpx.Response, url: str) -> dict:
    """安全解析 ASR 响应为 JSON。

    响应非 JSON（空响应 / HTML 错误页 / 网关 404）时，原来直接 .json() 会抛
    JSONDecodeError，把真实原因（HTTP 状态码 + 正文）掩盖掉。这里改成抛出
    带状态码与正文片段的 HTTPException，便于排查（如地址填成 compatible-mode/v1）。
    """
    try:
        return resp.json()
    except Exception:
        body = (resp.text or "").strip()[:300]
        raise HTTPException(
            500,
            f"ASR 接口返回非 JSON (HTTP {resp.status_code}) @ {url}：{body or '（空响应）'}。"
            "请检查语音接口地址是否为异步 ASR 网关（如 .../api/v1）及密钥是否正确。",
        )


async def _parse_voice(media_bytes: bytes, content_type: str) -> str:
    """语音转文字：阿里云异步 ASR 协议（提交任务 → 轮询 → 取结果）。

    接口地址动态：base = voice_base_url（为空直接用默认阿里云原生网关，不回退 ai_base_url），
    提交端点 = {base}/services/audio/asr/transcription，轮询 = {base}/tasks/{id}，
    子路径由代码自动拼接，配置只需填到 .../api/v1 这一层。
    密钥 = voice_api_key（为空回退 ai_api_key）。
    """
    ai = await get_ai_settings()
    # strip 防止配置/同步带入首尾空格（实测出现过 api_key 前导空格导致鉴权失败）
    # 注意：异步 ASR 走原生网关 /api/v1，与 ai_base_url（OpenAI 兼容的 chat 端点
    # .../compatible-mode/...）是两套服务，不能混用。voice_base_url 为空时应回退到
    # DEFAULT_ASR_BASE_URL，绝不能回退到 ai_base_url，否则会拼成 .../compatible-mode/.../asr 而 404。
    base = (ai.get("voice_base_url") or DEFAULT_ASR_BASE_URL).strip().rstrip("/")
    api_key = (ai.get("voice_api_key") or ai.get("ai_api_key") or "").strip()
    voice_model = (ai.get("voice_model") or "").strip()

    b64_audio = base64.b64encode(media_bytes).decode()
    data_uri = f"data:{_audio_mime(content_type)};base64,{b64_audio}"
    asr_headers = {
        "Authorization": f"Bearer {api_key}",
        "Content-Type": "application/json",
    }

    submit_url = f"{base}/services/audio/asr/transcription"
    async with httpx.AsyncClient(timeout=60) as http:
        submit = await http.post(
            submit_url,
            headers={**asr_headers, "X-DashScope-Async": "enable"},
            json={
                "model": voice_model,
                "input": {"file_urls": [data_uri]},
                "parameters": {"language_hints": ["zh", "en"]},
            },
            timeout=30,
        )
        submit_data = _asr_json(submit, submit_url)
        if submit.status_code not in (200, 201):
            raise HTTPException(500, f"提交识别任务失败 (HTTP {submit.status_code}): {submit_data.get('message', submit_data)}")

        task_id = submit_data.get("output", {}).get("task_id")
        if not task_id:
            raise HTTPException(500, f"未获取到 task_id: {submit_data}")

        for _ in range(30):
            await asyncio.sleep(1)
            poll = await http.get(
                f"{base}/tasks/{task_id}",
                headers=asr_headers,
                timeout=10,
            )
            poll_data = _asr_json(poll, f"{base}/tasks/{task_id}")
            status = poll_data.get("output", {}).get("task_status", "")
            if status == "SUCCEEDED":
                results = poll_data.get("output", {}).get("results", [])
                log.info("[media_parser] ASR SUCCEEDED results: %s", results)
                if not results:
                    return "（识别结果为空）"
                trans_url = results[0].get("transcription_url", "")
                if trans_url:
                    trans_resp = await http.get(trans_url, timeout=10)
                    trans_data = trans_resp.json()
                    log.info("[media_parser] transcription_url content: %s", str(trans_data)[:500])
                    transcripts = trans_data.get("transcripts", [])
                    text = transcripts[0].get("text", "") if transcripts else ""
                else:
                    text = results[0].get("transcription", "")
                return text or "（识别结果为空）"
            if status in ("FAILED", "CANCELLED"):
                raise HTTPException(500, f"识别任务失败: {poll_data.get('output', {}).get('message', status)}")

    raise HTTPException(500, "语音识别超时（30秒）")


async def _parse_visual(kind: str, media_bytes: bytes, content_type: str) -> str:
    b64 = base64.b64encode(media_bytes).decode()
    ct = content_type.lower()
    if "png" in ct:
        mime = "image/png"
    elif "webp" in ct:
        mime = "image/webp"
    else:
        mime = "image/jpeg"
    data_url = f"data:{mime};base64,{b64}"
    prompt = "请用中文简洁描述这张图片的内容。" if kind == "image" else "请用中文简洁描述这个视频截图的内容。"

    client, ai = await get_client_for("vision")
    resp_ai = await client.chat.completions.create(
        model=ai["vision_model"],
        messages=[
            {
                "role": "user",
                "content": [
                    {"type": "image_url", "image_url": {"url": data_url}},
                    {"type": "text", "text": prompt},
                ],
            }
        ],
        max_tokens=300,
    )
    return resp_ai.choices[0].message.content or ""