在DeepSeek-R1引发全球AI领域关注之际，其突破性的推理能力已通过多项测试得到验证：模型不仅将AIME数学竞赛准确率从15.6%提升至86.7%，更在Codeforces编程竞赛中超越96.3%人类参与者，展现出真实的数学直觉与迁移学习能力。然而作为纯文本模型，其官方版本存在多模态能力缺失与功能互斥的局限。
我选择通过Dify构建智能编排层：以DeepSeek-R1作为推理引擎，驱动更强大模型的多模态能力，实现文件解析与网络连接的协同运作。
在 Dify 中创建一个空白应用，选择 Chatflow 类型，打开工作区点击右上角的“功能”按钮，打开“文件上传”功能，勾选“文档”和“图片”类型。
按照上图编排工作流，核心思路就是解析文档和图片内容，交给 DeepSeek-R1 只生成推理内容，再把文档或图片以及解析到的内容和 R1 推理全部传给 Gemini 多模态模型，最终由 Gemini 来回答用户问题。
DeepSeek-R1 扮演“优等生”的角色，专注于问题分解和逻辑推理。其核心任务是输出完整的思考过程，而不是直接提供答案。
在编写系统提示时，建议编写结构化提示，例如使用 XML 格式，这可以增强模型对问题任务的分解。
<Role>You are an LLM with reasoning capabilities.Unlike other LLMs, you can output your complete thinking process.</Role><Task>Your task is to assist other LLMs that lack reasoning capabilities.You need to output complete thinking processes for other LLMs based on user questions.<Steps>"Step 1": "Receive questions from users.""Step 2": "Conduct deep reasoning and analysis on user questions.""Step 3": "Elaborate on the reasoning process and logic, ensuring the process is complete and easy to understand.""Step 4": "Output the complete reasoning process, no final answer needed."</Steps></Task><Limitations>Do not output the final answer, only output the thinking process.Do not explain your own capabilities or limitations.</Limitations>
In addition, we need to adjust the user input content, adding the content from the doc extractor:<User Query>{{Start}}</User Query><File>{{text}}</File>
Gemini 是一种具有强大视觉能力的多模态模型，依靠 R1 推理框架结合多模态数据并生成最终答案。其优势在于图像解析和结果优化。注意需要在此节点中启用LLM的视觉功能以获得解析图片和文档的能力。
<Role>You are an LLM that excels at learning.</Role><Task>You need to learn from others' thinking processes about problems, enhance your results with their thinking, and then provide your answer.<Steps>"Step 1": "Receive thinking process from DeepSeek-R1 model.""Step 2": "Carefully study and understand DeepSeek-R1's reasoning logic and steps.""Step 3": "Generate final answer based on DeepSeek-R1's thinking, combined with image capabilities.""Step 4": "Output the final answer, no need to explain the thinking process."</Steps></Task><Limitations>Do not repeat DeepSeek-R1's thinking process, only output the final answer.Do not explain your own capabilities or learning process.Ensure the answer is accurate and relevant to the question.</Limitations>
