crm_project/server/app/services/ocr_service.py

"""
OCR 服务 — 基于 3090 节点 Qwen3.5-27B (Vision)
对发票/名片图片做 AI 视觉理解，提取结构化数据。
"""
from __future__ import annotations

import base64
import json
import re
import httpx
from app.core.config import settings

INVOICE_PROMPT = """你是一个专业的发票OCR解析器。请分析图片中的发票/票据，提取以下结构化信息，以 JSON 格式返回：

{
  "merchant": "开票方/销售方名称",
  "amount": 金额数字(不带货币符号),
  "date": "YYYY-MM-DD 格式的开票日期",
  "invoice_code": "发票代码(如有)",
  "invoice_number": "发票号码(如有)",
  "tax_rate": "税率(如有)",
  "tax_amount": 税额数字(如有),
  "items": "发票上的商品/服务名称",
  "buyer": "购买方/抬头(如有)",
  "remark": "备注信息(如有)"
}

只输出 JSON，不需要解释。如果某个字段无法识别，设为 null。"""

BUSINESS_CARD_PROMPT = """你是一个名片OCR解析器。请分析图片中的名片，提取以下信息并以 JSON 返回：

{
  "name": "姓名",
  "company": "公司名称",
  "title": "职位",
  "phone": "电话号码",
  "email": "邮箱",
  "address": "地址",
  "other": "其他信息"
}

只输出 JSON。无法识别的字段设为 null。"""


async def ocr_image(
    image_base64: str,
    scene: str = "invoice",
) -> dict:
    """
    调用 3090 Qwen-VL 对图片做视觉理解/OCR。

    Args:
        image_base64: base64 编码的图片数据
        scene: "invoice" | "business_card" | "general"

    Returns:
        {"success": True, "data": {...提取的结构化数据...}}
    """
    fallback = {"success": False, "data": {}, "error": "OCR 服务不可用"}

    if not settings.OLLAMA_3090_BASE_URL:
        return fallback

    prompt = INVOICE_PROMPT if scene == "invoice" else (
        BUSINESS_CARD_PROMPT if scene == "business_card" else
        "请详细描述图片中的所有文字内容，以 JSON 格式输出。"
    )

    url = f"{settings.OLLAMA_3090_BASE_URL}/api/chat"
    payload = {
        "model": settings.OLLAMA_3090_MODEL,
        "messages": [
            {
                "role": "user",
                "content": "/no_think\n" + prompt,
                "images": [image_base64],  # Ollama vision 格式
            },
        ],
        "stream": False,
        "options": {
            "temperature": 0.1,
            "num_predict": 2000,
        },
    }

    try:
        async with httpx.AsyncClient(timeout=120.0) as client:
            resp = await client.post(url, json=payload)
            if resp.status_code != 200:
                print(f"[OCR] 3090 返回 {resp.status_code}: {resp.text[:200]}")
                return {"success": False, "data": {}, "error": f"VL 模型返回 {resp.status_code}"}

            data = resp.json()
            # Qwen3.5 的 CoT 推理放在 message.thinking，最终结果在 message.content
            content = data.get("message", {}).get("content", "")
            thinking = data.get("message", {}).get("thinking", "")

            # 优先从 content 提取 JSON，回退到 thinking
            for text_source in [content, thinking]:
                if not text_source:
                    continue
                cleaned = re.sub(r'<think>.*?</think>', '', text_source, flags=re.DOTALL).strip()
                json_match = re.search(r'\{[\s\S]*\}', cleaned)
                if json_match:
                    try:
                        result = json.loads(json_match.group())
                        print(f"[OCR] 解析成功: {list(result.keys())}")
                        return {"success": True, "data": result}
                    except json.JSONDecodeError:
                        continue

            # 没有提取到 JSON，返回原始文本
            raw = content or thinking
            print(f"[OCR] 未能提取 JSON, 内容长度: content={len(content)}, thinking={len(thinking)}")
            return {"success": True, "data": {"raw_text": raw[:2000]}}

    except httpx.TimeoutException:
        print("[OCR] 3090 超时（60s）")
        return {"success": False, "data": {}, "error": "VL 模型响应超时"}
    except json.JSONDecodeError as e:
        print(f"[OCR] JSON 解析失败: {e}")
        return {"success": False, "data": {}, "error": f"JSON 解析失败: {e}"}
    except Exception as e:
        print(f"[OCR] 错误: {e}")
        return {"success": False, "data": {}, "error": str(e)}


TEXT_INVOICE_PROMPT = """你是一个专业的发票数据提取器。以下是一份发票/票据的文本内容（来自 PDF 转换后的 Markdown 或纯文本）。
请从中提取以下结构化信息，以 JSON 格式返回：

{
  "merchant": "开票方/销售方名称",
  "amount": 金额数字(不带货币符号),
  "date": "YYYY-MM-DD 格式的开票日期",
  "invoice_code": "发票代码(如有)",
  "invoice_number": "发票号码(如有)",
  "tax_rate": "税率(如有)",
  "tax_amount": 税额数字(如有),
  "items": "发票上的商品/服务名称",
  "buyer": "购买方/抬头(如有)",
  "remark": "备注信息(如有)"
}

只输出 JSON，不需要解释。如果某个字段无法识别，设为 null。
注意：文本可能是从 PDF 转换而来，格式可能不规整，请智能识别。"""


async def extract_invoice_from_text(
    text: str,
    scene: str = "invoice",
) -> dict:
    """
    用 LLM 从纯文本（MD/TXT）中提取发票结构化数据。
    不走视觉模型，纯文本理解，更快更准。
    """
    fallback = {"success": False, "data": {}, "error": "AI 文本提取服务不可用"}

    if not settings.OLLAMA_3090_BASE_URL:
        return fallback

    prompt = TEXT_INVOICE_PROMPT if scene == "invoice" else (
        BUSINESS_CARD_PROMPT if scene == "business_card" else
        "请从以下文本中提取所有关键信息，以 JSON 格式输出。"
    )

    # 限制文本长度，避免 token 爆炸
    truncated = text[:8000] if len(text) > 8000 else text

    url = f"{settings.OLLAMA_3090_BASE_URL}/api/chat"
    payload = {
        "model": settings.OLLAMA_3090_MODEL,
        "messages": [
            {
                "role": "user",
                "content": f"/no_think\n{prompt}\n\n--- 以下是发票文本内容 ---\n\n{truncated}",
                # 不传 images —— 纯文本模式
            },
        ],
        "stream": False,
        "options": {
            "temperature": 0.1,
            "num_predict": 2000,
        },
    }

    try:
        async with httpx.AsyncClient(timeout=120.0) as client:
            resp = await client.post(url, json=payload)
            if resp.status_code != 200:
                print(f"[TextExtract] 3090 返回 {resp.status_code}: {resp.text[:200]}")
                return {"success": False, "data": {}, "error": f"LLM 返回 {resp.status_code}"}

            data = resp.json()
            content = data.get("message", {}).get("content", "")
            thinking = data.get("message", {}).get("thinking", "")

            for text_source in [content, thinking]:
                if not text_source:
                    continue
                cleaned = re.sub(r'<think>.*?</think>', '', text_source, flags=re.DOTALL).strip()
                json_match = re.search(r'\{[\s\S]*\}', cleaned)
                if json_match:
                    try:
                        result = json.loads(json_match.group())
                        print(f"[TextExtract] AI 提取成功: {list(result.keys())}")
                        return {"success": True, "data": result}
                    except json.JSONDecodeError:
                        continue

            raw = content or thinking
            print(f"[TextExtract] 未能提取 JSON, 内容: {raw[:200]}")
            return {"success": True, "data": {"raw_text": raw[:2000]}}

    except httpx.TimeoutException:
        print("[TextExtract] 3090 超时")
        return {"success": False, "data": {}, "error": "LLM 响应超时"}
    except Exception as e:
        print(f"[TextExtract] 错误: {e}")
        return {"success": False, "data": {}, "error": str(e)}