让AI读懂PPT图表！RAG系统从60分到95分的进化之路，LlamaParse+多模态实战全解析 原创

根据提供的资源数据，2017年大数据开发职位的招聘人数大幅增加。具体来说，2016年大数据开发的招聘人数为5,667人，而2017年则增加到41,831人。因此，2017年大数据开发职位比2016年增加了36,164人 (41,831 - 5,667 = 36,164)。

os.environ["LLAMA_CLOUD_API_KEY"] = "xxx"

parser = LlamaParse(
        result_type="json",
        use_vendor_multimodal_model=True,
        vendor_multimodal_model_name="gemini-2.0-flash-001",
        language="ch_sim"
    )
md_result = parser.get_json_result(file_path)
doc_id = md_result[0]["job_id"]
pages = md_result[0]["pages"]
# 下载图片到本地
parser.get_images(md_result, download_path="data_images")

dataset = []
docs = []
base64_map = {}
for page in pages:
    md = page["md"]
    page_number = page["page"]

    # 查找并上传对应页码的图片
    local_image_path = find_image_by_page(
        "data_images", doc_id, page_number)
    with open(local_image_path, "rb") as image_file:
        image_base64 = base64.b64encode(image_file.read()).decode('utf-8')

    # 添加到dataset
    dataset.append({
        'content': md,
        'image_base64': image_base64,
        'page_number': page_number
    })
    docs.append(
        Document(
            page_cnotallow=md,
            metadata={
                "page_number": page_number,
            })
    )
    base64_map[page_number] = image_base64
    
vectorstore = Milvus.from_documents(
    documents=docs,
    embedding=embeddings,
    connection_args={"host": "127.0.0.1", "port": "19530"},
    drop_old=True,  # Drop the old Milvus collection if it exists
    collection_name="collection_ppt",
)

dat = vectorstore.similarity_search(query=question, k=5)

image_base64_list = []
chunk = []
for doc in dat:
    page_number = doc.metadata["page_number"]
    print(page_number)
    image_base64 = base64_map.get(page_number)
    if image_base64:
        image_base64_list.append(image_base64)
        chunk.append(doc.page_content)
        
openai_api_key = os.environ["OPENAI_API_KEY"]  # 替换为你的 OpenAI API Key
headers = {
    "Content-Type": "application/json",
    "Authorization": f"Bearer {openai_api_key}"
}

# 构建 messages 内容
messages_content = [
    {"type": "text", "text": """The following is the Markdown text and image information parsed in the slide. Markdown text has attempted to convert the relevant charts into tables. Give priority to using picture information to answer questions. Use Markdown text information only when the image cannot be understood.
 
Here is the context:
---------
{context}
---------

Question: {question}
""".format(cnotallow="\n".join(chunk), questinotallow=question)}
]

if image_base64_list:
    # 添加所有检索到的图片
    for img_base64 in image_base64_list:
        messages_content.append({
            "type": "image_url",
            "image_url": {
                "url": f"data:image/jpeg;base64,{img_base64}"
            }
        })

payload = {
    "model": "gpt-4.1",
    "messages": [
        {
            "role": "user",
            "content": messages_content
        }
    ],
    "temperature": 0.1,
    "max_tokens": 1000# 根据需要调整
}

# 发送请求到 OpenAI API
try:
    response = requests.post(
        "https://api.openai-proxy.com/v1/chat/completions",
        headers=headers,
        jsnotallow=payload
    )
    response.raise_for_status()  # 检查请求是否成功

    # 解析并打印结果
    result = response.json()
    print("OpenAI 分析结果:")
    print(result["choices"][0]["message"]["content"])

except requests.exceptions.RequestException as e:
    print(f"请求 OpenAI API 失败: {e}")
    if hasattr(e, 'response') and e.response:
        print(f"错误详情: {e.response.text}")