Video_cutter/utils/llm_director.py

import os
import re
import json
import httpx
from volcenginesdkarkruntime import Ark

client = Ark(
    base_url="https://ark.cn-beijing.volces.com/api/v3",
    api_key="817dff39-5586-4f9b-acba-55004167c0b1",
    timeout=1800
)

def read_input(json_file):
    with open(json_file, 'r', encoding='utf-8') as file:
        data = json.load(file)

    # 将数据转换为字符串
    json_string = json.dumps(data, ensure_ascii=False, indent=4)
    return json_string

def convert_json_to_dict(json_string, output_file_path=None):
    match = re.search(r'\{.*\}', json_string, re.DOTALL)
    if match:
        json_content = match.group(0)  # 提取匹配的内容
    else:
        print("错误: 未找到有效的 JSON 内容")
        return {}

    try:
        # 尝试将提取的内容转换为字典
        data_dict =  json.loads(json_content)
        if output_file_path:
            with open(output_file_path, 'w', encoding='utf-8') as file:
                json.dump(data_dict, file, ensure_ascii=False, indent=4)
        return data_dict
       
    except json.JSONDecodeError as e:
        print(f"JSON 解码错误: {e}")
        return {}
    

def caption_correct(user_prompt):

    count = len(user_prompt)
    cut = "不需要断句，直接输出纠正错别字后的文本"
    if count > 15:
        cut = """必须要断句，且只能断句一处；用符号"-"进行断句，断句位置要合理"""

    system_prompt = f"""
            ## 对输入文本进行错别字纠正，错别字通常都是因为字词的发音相似而引起的。如"歌莉娅"写成了"哥李呀"；"的"写成了"得"。
            ## 用户输入：{user_prompt}
            ## 要求：
            - 需要去除文本中的标点符号。
            - {cut}
            - 与"歌莉娅"发音相似的词，都要纠正为"歌莉娅"
            - 只输出纠正错别字后的文本，不能有任何多余的文本输出
        """

    completion = client.chat.completions.create(
        messages = [
            {"role": "system", "content": system_prompt},
        ],
        model="deepseek-v3-241226", # ep-20241018084532-cgm84 deepseek-v3-241226 deepseek-r1-250120
        temperature = 0.01,
        max_tokens = 500
    )

    result = completion.choices[0].message.content.replace("-", "\n")
    return result



def director(user_prompt):

    system_prompt = """
            ## 我需要剪辑一个衣服口播讲解视频，该视频需要包含三个视频片段；请从用户输入中挑选讲解衣服面料、版型、工艺的视频片段各一到三个；并输出三个衣服口播讲解视频脚本（各视频片段组合方式）；并以JSON格式进行输出。
            ## 输出案例：
            ```json
            {
                "面料":["clip_001.mp4", "clip_004.mp4", "clip_006.mp4"],
                "版型":["clip_011.mp4", "clip_002.mp4", "clip_007.mp4"],
                "工艺":["clip_012.mp4", "clip_013.mp4", "clip_014.mp4"],
                "脚本":[["clip_001.mp4", "clip_011.mp4", "clip_013.mp4"], ["clip_004.mp4", "clip_014.mp4", "clip_002.mp4"], ["clip_006.mp4", "clip_011.mp4", "clip_012.mp4"]]
            }
            ## 严格按照输出案例的格式输出结果，不能输出任何多余的内容。
            ## 尽可能找句子比较长的视频片段
        """

    completion = client.chat.completions.create(
        messages = [
            {"role": "system", "content": system_prompt},
            {"role": "user", "content": user_prompt},
        ],
        model="deepseek-r1-250120", # ep-20241018084532-cgm84 deepseek-v3-241226 deepseek-r1-250120
        temperature = 0.01,
        max_tokens = 500
    )
    return completion.choices[0].message.content

def director_json(json_path):
    user_input = read_input(json_path)
    answer = director(user_input)
    output_path = json_path.replace("filter_4", "script")
    dict_answer = convert_json_to_dict(answer, output_path)
    return dict_answer




if __name__ == "__main__":

    
    print(director_json("output/filter_4/videoa.json"))