说明

只是为了速记一下这个实践过程。整体上说，这个结果并不是那么好用，但有一些可以借鉴的地方。

先看结果：

生成的PPT

说的直白点，就是用大模型生成了一堆没太有意义的文字，然后做成ppt。所以实用是不成的，但是里面有一些过程可以借鉴。

内容

1 项目地址

ppt_generator

整个项目没有几个文件，感觉也就是一个原型实验。

需要的环境是本地ollama(恰好我有），然后拉一个模型

项目里还漏了 requirement.txt，我的环境下，主要再安装一个包就可以了

pip3 install python-pptx  -i https://mirrors.aliyun.com/pypi/simple/

2 使用

启动 streamlit run main.py,然后会弹出一个web页面，输入一个主题，然后就返一个ppt给你。整体过程大概1~2分钟，考虑是笔记本，而且也看到什么资源占用，所以可以认为是不太占资源的。

3 核心代码

使用langchain加载了ollama模型，生成数据

def slide_data_gen(topic):
    llm = Ollama(model="dolphin2.1-mistral",
                 temperature="0.4")

    slide_data = []

    point_count = 5

    slide_data.append(extract_items(llm(f"""
    You are a text summarization and formatting specialized model that fetches relevant information

    For the topic "{topic}" suggest a presentation title and a presentation subtitle it should be returned in the format :
    << "title" | "subtitle >>

    example :
    << "Ethics in Design" | "Integrating Ethics into Design Processes" >>
    """)))

    slide_data.append(extract_items(llm(f"""
    You are a text summarization and formatting specialized model that fetches relevant information
            
    For the presentation titled "{slide_data[0][0]}" and with subtitle "{slide_data[0][1]}" for the topic "{topic}"
    Write a table of contents containing the title of each slide for a 7 slide presentation
    It should be of the format :
    << "slide1" | "slide2" | "slide3" | ... | >>
            
    example :
    << "Introduction to Design Ethics" | "User-Centered Design" | "Transparency and Honesty" | "Data Privacy and Security" | "Accessibility and Inclusion" | "Social Impact and Sustainability" | "Ethical AI and Automation" | "Collaboration and Professional Ethics" >>          
    """)))

    for subtopic in slide_data[1]:

        data_to_clean = llm(f"""
        You are a content generation specialized model that fetches relevant information and presents it in clear concise manner
                
        For the presentation titled "{slide_data[0][0]}" and with subtitle "{slide_data[0][1]}" for the topic "{topic}"
        Write the contents for a slide with the subtopic {subtopic}
        Write {point_count} points. Each point 10 words maximum.
        Make the points short, concise and to the point.
        """)

        cleaned_data = llm(f"""
        You are a text summarization and formatting specialized model that fetches relevant information and formats it into user specified formats
        Given below is a text draft for a presentation slide containing {point_count} points , extract the {point_count} sentences and format it as :
                    
        << "point1" | "point2" | "point3" | ... | >>
                    
        example :
        << "Foster a collaborative and inclusive work environment." | "Respect intellectual property rights and avoid plagiarism." | "Uphold professional standards and codes of ethics." | "Be open to feedback and continuous learning." >>

        -- Beginning of the text --
        {data_to_clean}
        -- End of the text --         
        """)

        slide_data.append([subtopic] + extract_items(cleaned_data))

    return slide_data

根据数据生成ppt

def ppt_gen(slide_data):
    ppt = Presentation()

    # Setting Background
    slide_master = ppt.slide_master
    slide_master.background.fill.solid()
    slide_master.background.fill.fore_color.rgb = RGBColor(0, 0, 0)

    # Title Screen
    curr_slide = ppt.slides.add_slide(ppt.slide_layouts[0])
    curr_slide.shapes.title.text = slide_data[0][0]
    curr_slide.shapes.title.text_frame.auto_size = MSO_AUTO_SIZE.TEXT_TO_FIT_SHAPE
    curr_slide.shapes.title.text_frame.paragraphs[0].runs[0].font.color.rgb = RGBColor(
        255, 255, 255)
    curr_slide.shapes.placeholders[1].text = slide_data[0][1]
    curr_slide.shapes.placeholders[1].text_frame.auto_size = MSO_AUTO_SIZE.TEXT_TO_FIT_SHAPE
    curr_slide.shapes.placeholders[1].text_frame.paragraphs[0].runs[0].font.color.rgb = RGBColor(
        255, 255, 255)

    # Overview
    curr_slide = ppt.slides.add_slide(ppt.slide_layouts[1])
    curr_slide.shapes.title.text = "Overview"
    curr_slide.shapes.title.text_frame.auto_size = MSO_AUTO_SIZE.TEXT_TO_FIT_SHAPE
    curr_slide.shapes.title.text_frame.paragraphs[0].runs[0].font.color.rgb = RGBColor(
        255, 255, 255)
    for content in slide_data[1]:
        tframe = curr_slide.shapes.placeholders[1].text_frame
        tframe.auto_size = MSO_AUTO_SIZE.TEXT_TO_FIT_SHAPE
        para = tframe.add_paragraph()
        para.text = content
        para.level = 1
        para.font.color.rgb = RGBColor(
            255, 255, 255)

    # Content Slides
    for curr_slide_data in slide_data[2:]:
        curr_slide = ppt.slides.add_slide(ppt.slide_layouts[1])
        curr_slide.shapes.title.text = curr_slide_data[0]
        curr_slide.shapes.title.text_frame.auto_size = MSO_AUTO_SIZE.TEXT_TO_FIT_SHAPE
        curr_slide.shapes.title.text_frame.paragraphs[0].font.color.rgb = RGBColor(
            255, 255, 255)
        for content in curr_slide_data[1:]:
            tframe = curr_slide.shapes.placeholders[1].text_frame
            tframe.auto_size = MSO_AUTO_SIZE.TEXT_TO_FIT_SHAPE
            para = tframe.add_paragraph()
            para.text = content
            para.level = 1
            para.font.color.rgb = RGBColor(
                255, 255, 255)

    # Thank You Screen
    curr_slide = ppt.slides.add_slide(ppt.slide_layouts[2])
    curr_slide.shapes.placeholders[1].text = "Thank You"

    curr_slide.shapes.placeholders[1].text_frame.paragraphs[0].font.color.rgb = RGBColor(
        255, 255, 255)
    curr_slide.shapes.placeholders[1].text_frame.paragraphs[0].font.size = Pt(
        96)
    curr_slide.shapes.placeholders[1].text_frame.paragraphs[0].alignment = PP_ALIGN.CENTER

    # f"{sanitize_string(slide_data[0][0])}.pptx"
    ppt_stream = io.BytesIO()
    ppt.save(ppt_stream)
    ppt_stream.seek(0)

    return ppt_stream

4 总结

• 1 可以参考langchain，进行增强。langchain本身还有agent的能力，继续融合后，是有可能完全不一样的。例如当你提到一个论点，agent可以自动查询实时数据，找到论据来support.
• 2 ppt生成（及读取）。之前比较少用python操作ppt,现在看来，读和简单写应该是没问题的。