Home›Practice
Practice
Practice
Hands-on guides and first-hand notes on agent frameworks, skills and AI tools.
Total 217
Deploying MiniMax-M3 with vLLMServe the open-source MiniMax-M3 model with vLLM on 4 or 8 GPUs, then call it through an OpenAI-compatible API for chat and image input.Deploying Kimi K2.5 with SGLangRun Kimi K2.5 on eight 96GB GPUs with SGLang, expose an OpenAI-compatible API, and test chat, streaming and a weather tool call.Serving Tencent Hunyuan-A13B-Instruct with vLLMRun Tencent's Hunyuan-A13B-Instruct on four GPUs with vLLM, then call it through an OpenAI-compatible API with thinking output and a weather tool call.Serving GLM-4.5-Air with vLLMDownload GLM-4.5-Air from ModelScope, run it offline with vLLM's Python API, then serve it on four GPUs with an OpenAI-compatible endpoint and reasoning output.Deploying DeepSeek-V4-Flash with SGLangServe DeepSeek-V4-Flash through SGLang's OpenAI-compatible API: the official Docker image on two RTX PRO 6000 cards, or a Python install on four H200 cards.Deploying DeepSeek-V4-Flash with vLLMServe DeepSeek-V4-Flash on a four-GPU RTX PRO 6000 machine behind an OpenAI-compatible API, using the official vLLM Docker image or a pinned SM120 source build.Fine-tuning large language models: a practical startFine-tuning retrains a model on your own examples. Learn when it is worth doing, why to set a baseline first, and how to choose a technique, base model and data format on Microsoft Foundry.The generative AI application lifecycle (LLMOps)Learn the generative AI application lifecycle: LLMOps metrics, the ideate, build and operationalize stages, and how Microsoft Foundry and PromptFlow support each step.Securing generative AI applicationsLearn the main risks to generative AI applications, including data poisoning and prompt injection, and the defenses that help: data protection, output validation, security testing and red teaming.Designing user experience for AI applicationsDesigning AI apps means making them useful, reliable, accessible and pleasant, and building appropriate trust through explanations, user control and feedback for when the AI gets things wrong.Building low-code AI applicationsLearn how Power Platform's Copilot, AI Builder and Copilot Studio let you build apps, automated flows and AI agents with little or no code, and where each one fits.Building Image Generation ApplicationsBuild a Python image generation app with a gpt-image model: decode the base64 result, save it as a PNG, edit images with a mask, and constrain output with a metaprompt.Building Text Generation Applications with LLM APIsBuild a Python text generation app with the openai library: keep the key in an environment variable, call the Responses API, then tune results with prompts, output length and temperature.Generated Knowledge PromptingGenerated knowledge prompting has a model write background facts about a question first, then uses those facts in a second prompt to reach an answer. Here are both steps and their limits.Reflexion: Self-Reflective Prompting for Language AgentsReflexion lets a language agent turn feedback from each attempt into written self-reflections, store them in memory and use them to improve the next try, without fine-tuning the underlying model.Tree of Thoughts (ToT) PromptingTree of Thoughts (ToT) prompting has a model generate and judge intermediate reasoning steps, then search the options with backtracking. Here is how the idea works and how to try a one-prompt version.Self-Consistency PromptingSample several reasoning paths with few-shot chain-of-thought prompting, then pick the most consistent answer. Follow the arithmetic example to run the self-consistency workflow.Program-Aided Language Models (PAL): Let the Model Write Code to ComputeTurn a date question into Python code with an LLM, then let an interpreter compute the answer. Follow the example to run a small program-aided prompting app.Getting Started with Dify Plugin DevelopmentSet up the Dify plugin CLI, generate a Python plugin project, configure remote debugging in the .env file, and run your first plugin locally.Embedding a Dify Chatbot in Your WebsiteEmbed a published Dify web app in your website as a chat bubble or an iframe, with the token, configuration, styling options and troubleshooting steps.Publishing a Dify App as an MCP ServerTurn on the MCP service for a Dify app, get its server URL, and connect it to Claude Desktop or Cursor so those tools can call your app directly.Dify Code and HTTP Request Nodes: Custom Logic and External APIsDify's code node runs Python or JavaScript for custom logic, and the HTTP request node calls external APIs. This guide covers writing code node functions, output limits, sandbox rules, and configuring HTTP requests.Dify LLM Node: Prompts, Parameters and Structured OutputDify's LLM node sends a configured prompt to a language model and returns the reply. This guide covers parameters, prompt variables, knowledge base context, structured output, memory, reasoning output and retries.Building a Dify Agent App: Prompts, Variables and ToolsA Dify Agent is a chat app in which the model decides when to use tools to answer. This guide covers writing the prompt, adding variables and tools, grounding answers in a knowledge base, and debugging and publishing.