{"id":4127,"date":"2025-02-05T12:00:53","date_gmt":"2025-02-05T20:00:53","guid":{"rendered":"https:\/\/devblogs.microsoft.com\/semantic-kernel\/?p=4127"},"modified":"2025-02-05T12:02:19","modified_gmt":"2025-02-05T20:02:19","slug":"guest-blog-let-your-copilot-declarative-agent-think-deep-with-deepseek-r1","status":"publish","type":"post","link":"https:\/\/devblogs.microsoft.com\/agent-framework\/guest-blog-let-your-copilot-declarative-agent-think-deep-with-deepseek-r1\/","title":{"rendered":"Guest Blog: Let your Copilot Declarative Agent think deep with DeepSeek-R1"},"content":{"rendered":"

Today we’d like to feature a guest author on our Semantic Kernel blog, Mahmoud Hassan, a Microsoft Valuable Professional (MVP) focused on AI. We’ll turn it over to him to dive into his work below.<\/p>\n

 <\/p>\n

In recent days, there has been significant attention in the AI community regarding DeepSeek-R1 and its capabilities.<\/p>\n

Many people are playing with it. For instance, Fabian Williams<\/a> yesterday shared his experiment here: https:\/\/lnkd.in\/dgZ8hjgB<\/a> of running it locally. I thought, maybe today is my turn!<\/p>\n

However, I also thought it is an excellent opportunity to show a plugin design pattern I previously shared with you \u201cThe Intelligent Plugin Design Pattern\u201d<\/p>\n

\u2728 The Intelligent Plugin Design Pattern\nA design pattern that makes the Microsoft 365 Copilot API plugins more intelligent by giving it access to the intelligence of A large language model (LLM).<\/p>\n

\ud83d\udd17 More information here: https:\/\/lnkd.in\/dSBTzh2y<\/a><\/p>\n

\ud83d\udd25 The Think Deep Declarative Agent with DeepSeek-R1\nTo summarize, I created a simple declarative agent with one action, “Think Deep Action”, which connects via Dev Tunnel to my local workstation and call the “Think Deep API” powered by Semantic Kernel to access Ollama’s local deployment of DeepSeek-R1.<\/p>\n

The Agent Architecture:<\/strong><\/h3>\n

\"Image<\/a><\/p>\n

The Semantic Kernel Code:<\/h3>\n
using var ollamaClient = new OllamaApiClient(\r\n            uriString: \"http:\/\/localhost:11434\",\r\n            defaultModel: \"deepseek-r1:1.5b\");\r\nvar chatService = ollamaClient.AsChatCompletionService();\r\nChatMessageContent replyMessage = await chatService.GetChatMessageContentAsync(userQuery);\r\nreturn replyMessage.ToString();<\/code><\/pre>\n
\u2757I also tested another architecture for hosting the DeepSeek-R1 on Azure AI Foundry (Available since Jan 29, 2025<\/strong>). Just 3 lines of code, and we’re in the cloud. I \u2764\ufe0f #SemanticKernel<\/a> & #AzureAIFoundry<\/a> \"Image<\/a><\/p>\n
The Semantic Kernel Code:<\/h3>\n
var azureAIchatService = new ChatCompletionsClient(\r\n                endpoint: new Uri(\"https:\/\/DeepSeek-R1-ochnw.eastus.models.ai.azure.com\"),\r\n                credential: new Azure.AzureKeyCredential(\"<Add Key>\"))\r\n            .AsChatClient(\"DeepSeek-R1\")\r\n            .AsChatCompletionService();\r\nChatMessageContent replyMessage = await azureAIchatService.GetChatMessageContentAsync(userQuery);\r\nreturn replyMessage.ToString();\r\n<\/code><\/pre>\n
\ud83d\udca1[Bonus] The API plugin function states (Reasoning & responding)\nThe API plugin Function states is a hidden gem of the API plugin:\n\u2728 The Reasoning instructions let you control the orchestrator’s actions when calling your function.\n\u2728 The Responding instructions control how the orchestrator processes your function response and presents it to the user.<\/p>\n
Documentation Link: API plugin manifest schema 2.2 for Microsoft 365 Copilot<\/a><\/p>\n
In the bottom of the LinkedIn post below, there is an embedded video, which ends with a tutorial on using the Responding instruction to remove the <think \/> section of DeepSeek-R1.<\/p>\n