Home›Practice
Practice
Practice
Hands-on guides and first-hand notes on agent frameworks, skills and AI tools.
Total 217
Dify Workflow and Chatflow: Which Agentic App Type to BuildDify builds agentic workflows as either Workflow or Chatflow apps, both made by connecting nodes on a visual canvas. This guide explains how the two differ and how to choose between them.Dify Knowledge Base Retrieval Settings: How to Tune ThemDify knowledge base retrieval settings control how chunks are found and ranked. This guide covers the models, Top K and score threshold, the Fast, Deep and Research depths, and how to test and evaluate retrieval.Creating a knowledge base in DifyA knowledge base lets an AI app answer from your own documents. Learn how to create one in Dify from a data source or uploaded files, set sync, and test retrieval before connecting it to apps.Self-hosting Dify with Docker ComposeRun Dify on your own machine with Docker Compose. Learn the hardware and software requirements, startup and access, a sample config change, and the security and upgrade steps.Deploying LangChain agents to productionA working agent still needs a hosting model. Learn the LangSmith Cloud route from a GitHub repository, plus JavaScript framework hosting and the other LangSmith options.Observability and debugging for LangChain agentsTo debug an agent you need to see which tools it called, which prompts it generated and where it made each decision. Learn how LangSmith tracing records every step of a LangChain agent.Context engineering for agents with LangChainMost agent failures come from missing context rather than weak models. Learn the three context layers LangChain describes and how to control them with tools and middleware.Getting started with llama.cppGetting started with llama.cpp: install scripts to CMake source builds, one-command GGUF model runs, a built-in OpenAI-compatible server with web UI, and the compile flags for CUDA, Metal, and other backends.vLLM quickstartThe official vLLM quickstart: installing with uv and precompiled kernels, offline batched inference through generate and chat, online serving over the OpenAI protocol, and choosing attention backends.Serving Qwen3-VL with vLLMDeploying Qwen3-VL-4B-Instruct on vLLM and testing its multimodal range: a single RTX 5090 baseline, OpenAI-style image messages, offline video inference, and the VRAM-versus-truncation tradeoffs of vision workloads.Gemma 4 with Ollama and Open WebUIRunning gemma4:e4b with Ollama and building a local chat interface with Open WebUI: two Linux install paths, the port and mirror-source gotchas, and the auto-detection handshake in the browser.Serving MiniMax-M2 with vLLMDeploying MiniMax-M2 on vLLM: the 220GB weight budget, 4-GPU versus 8-GPU parallel setups, the minimax_m2 tool-call and reasoning parsers, and a tool-calling test built for agent workloads.Serving gpt-oss-20b with vLLMDeploying OpenAI's open gpt-oss-20b with vLLM: matching the pre-release vLLM wheel and FlashInfer, calling the model through transformers and vLLM, and tuning reasoning effort with a system prompt.Serving Kimi K2.5 with vLLMDeploy Moonshot's Kimi K2.5 on vLLM behind eight GPUs: thinking mode with official parameters, kimi_k2 reasoning and tool-call parsers, and a function-calling test that fixes two classic reasoning-model mistakes.Serving GLM-4.7-Flash with vLLMDeploy GLM-4.7-Flash with vLLM: enable thinking mode with the official sampling parameters, parse reasoning with --reasoning-parser, and verify its tool calling end to end with a small function-calling demo.DeepSeek-R1-Distill-Qwen with LangChainWrap a locally deployed DeepSeek-R1-Distill-Qwen-7B as a LangChain LLM class, then call it like any other LangChain model and split the thinking content from the final answer with a regex.Serving DeepSeek-R1-Distill-Qwen with vLLMServe DeepSeek-R1-Distill-Qwen-7B with vLLM: install pinned dependencies, download the model via ModelScope, run offline inference with the right thinking parameters, launch an OpenAI-compatible server, and benchmark throughput.How tools work in AI agents: from function to tool callTools are the functions an agent gives its LLM so it can act. After this guide you can write a typed tool, generate its description automatically, and follow how a tool call is made and run.What is an AI agent? The think, act, observe loopAn AI agent uses a language model as its reasoning core and tools to act on its environment. After this guide you can separate the model from the tools and explain the think, act, observe loop.Deploying scalable agents with Microsoft FoundryMoving an agent from a laptop to production means building the operational layer around the model. After this guide you can choose a deployment pattern and add evaluation gates and smoke tests to a release.Building a browser-operating agent with Browser-Use and PlaywrightBrowser-Use lets an AI agent open web pages, read them and act like a person. After this guide you can build a small browser agent with Playwright and know which actions need user approval.Getting started with Microsoft Agent FrameworkMicrosoft Agent Framework is Microsoft's unified framework for building AI agents. After this guide you can create an agent, run it, add tools and threads, and build a workflow.MCP in Action: Real-World Case StudiesEight real MCP cases, from multi-agent travel planning to a social-publishing server. The publishing case is the most useful for anyone building a write-capable MCP server, so it gets the most space here.Building a Semantic Product Search Service with MCPBuild a semantic product search as an MCP tool. Azure OpenAI embeddings go into PostgreSQL with pgvector and are queried by cosine distance, with hybrid search, recommendations, and caching.