from langchain.chat_models import init_chat_model
from langchain_core.messages import HumanMessage

model = init_chat_model(
    "your-multimodal-model-id",  # 示例：任意支持多模态的模型
    output_version="v1",
)

def build_multimodal_message(image_base64: str) -> HumanMessage:
    return HumanMessage(
        content=[
            {
                "type": "text",
                "text": "请根据下面这张图，判断近 30 天销售是否有明显异常，并用中文给出结论和可能原因。",
            },
            {
                "type": "image",
                "base64": image_base64,
                "mime_type": "image/png",
            },
        ]
    )

import base64
import os
from pathlib import Path

from langchain_ollama import ChatOllama
from langchain_deepseek import ChatDeepSeek
from langchain_core.messages import HumanMessage

def load_image_as_base64(path: str | Path) -> str:
    p = Path(path)
    with p.open("rb") as f:
        return base64.b64encode(f.read()).decode("utf-8")

def ocr_image_with_ollama(path: str | Path) -> str:
    image_b64 = load_image_as_base64(path)

    base_url = os.getenv("OLLAMA_BASE_URL", "http://localhost:11434")
    ocr_model_id = os.getenv("OCR_MODEL", "deepseek-ocr")
    prompt = os.getenv(
        "OCR_PROMPT",
        "请逐行识别图片中的文字，保持原有行结构，只输出纯文本，不要额外解释。",
    )

    ocr_model = ChatOllama(
        model=ocr_model_id,
        base_url=base_url,
        temperature=0,
        output_version="v1",
    )

    msg = HumanMessage(
        content=[
            {"type": "text", "text": prompt},
            {
                "type": "image",
                "base64": image_b64,
                "mime_type": "image/png",
            },
        ]
    )

    resp = ocr_model.invoke([msg])
    # 简单起见，这里直接用 text 字段；也可以从 resp.content_blocks 里抽取 text 块
    return resp.text.strip()

llm = ChatDeepSeek(
    model=os.getenv("DEEPSEEK_MODEL", "deepseek-reasoner"),
    api_key=os.getenv("DEEPSEEK_API_KEY"),
    api_base=os.getenv("DEEPSEEK_BASE_URL", "https://api.deepseek.com"),
    temperature=0.2,
    output_version="v1",
)

def build_message_from_ocr(image_path: str | Path, user_note: str) -> HumanMessage:
    ocr_text = ocr_image_with_ollama(image_path)

    system_text = (
        "你是一名电商运营分析师，擅长解读 GMV 等业务指标的趋势。\n"
        "给你一张报表截图和用户补充说明，请用中文输出：\n"
        "1）整体趋势概述；2）明显异常点；3）可能原因；4）后续建议。"
    )

    return HumanMessage(
        content=[
            {"type": "text", "text": system_text},
            {"type": "text", "text": f"用户补充说明：{user_note}"},
            {
                "type": "text",
                "text": f"截图 OCR 识别结果：\n{ocr_text}",
            },
        ]
    )

def demo_ocr_pipeline():
    msg = build_message_from_ocr("examples/gmv_trend.png", "最近 7 天做了两次大促")
    resp = llm.invoke([msg])
    print("标准 content_blocks：", resp.content_blocks)

response = model.invoke("画一张简单的增长曲线图，并简单解释含义。")
print(response.content_blocks)

# 可能的输出结构（示意）
# [
#   {"type": "text", "text": "这张图展示了一个稳定增长的趋势……"},
#   {"type": "image", "base64": "...", "mime_type": "image/jpeg"},
# ]

from typing import List, Dict, Any

def split_multimodal_output(blocks: List[Dict[str, Any]]):
    """把 content_blocks 拆成文本列表 + 图片列表，便于分别处理。"""
    texts = []
    images = []

    for block in blocks:
        if block.get("type") == "text":
            texts.append(block.get("text") or "")
        elif block.get("type") == "image":
            images.append(block)

    return texts, images

def demo_multimodal_output():
    response = model.invoke("画一张简单的增长曲线图，并简单解释含义。")
    texts, images = split_multimodal_output(response.content_blocks)

    print("文本部分：")
    print("\n".join(texts))

    print("\n图片数量：", len(images))
    for idx, img in enumerate(images):
        print(f"第 {idx+1} 张图片 mime_type={img.get('mime_type')}, "
              f"base64 前 50 字符={img.get('base64', '')[:50]}...")

from langchain_core.messages import HumanMessage

def build_report_analysis_message(
    image_path: str,
    user_note: str,
) -> HumanMessage:
    """构造报表截图分析的消息（图片先经过 deepseek-ocr 识别）。"""
    system_text = (
        "你是一名电商运营分析师，擅长解读 GMV 等业务指标的趋势。\n"
        "给你一张报表截图和用户补充说明，请用中文输出：\n"
        "1）整体趋势概述；2）明显异常点；3）可能原因；4）后续建议。"
    )

    ocr_text = ocr_image_with_ollama(image_path)

    return HumanMessage(
        content=[
            {"type": "text", "text": system_text},
            {"type": "text", "text": f"用户补充说明：{user_note}"},
            {"type": "text", "text": f"截图 OCR 识别结果：\n{ocr_text}"},
        ]
    )

def compose_text_from_blocks(blocks):
    return "\n".join(
        block["text"]
        for block in blocks
        if block.get("type") == "text" and block.get("text")
    )

def analyze_report(image_path: str, user_note: str) -> str:
    message = build_report_analysis_message(image_path, user_note)

    response = model.invoke([message])
    return compose_text_from_blocks(response.content_blocks)

{
    "type": "image",
    "base64": "...",
    "mime_type": "image/png",
    "purpose": "main_chart",  # 自定义字段
}