12 MCP, RAG, and Agents Cheat Sheets for AI Engineers

...explained with visuals.

​The Web MCP is here with 5,000 Monthly Credits​

Bright Data has launched a free tier of ​The Web MCP, the first and only MCP designed to give LLMs and autonomous agents unblocked, real-time access to the web.

Now you can: /scrape /search /crawl /navigate The live web with 5,000 free monthly credits.

Built for developers and researchers working with open-source tools:

Key features:

• Integrates seamlessly with your workflow, integrates with LangChain, AutoGPT, OpenAgents, and custom stacks.

• Enables agents to dynamically expand their context with live web data

• All major LLMs and IDEs are supported (locally hosted, SSE, and Streamable HTTP)

• No setup fees, no credit card required.

Whether you're building agentic workflows, RAG pipelines, or real-time assistants, The Web MCP is the protocol layer that connects your models to the open web.

​Start building with 5,000 free monthly credits here →​

Thanks to Bright Data for partnering today!

---

12 MCP, RAG, and Agents Cheat Sheets for AI Engineers

Here’s a recap of several visual summaries posted in the Daily Dose of Data Science newsletter.

​1) Function calling & MCP for LLMs:​

• Before MCPs became popular, AI workflows relied on traditional Function Calling for tool access.

• Now, MCP (Model Context Protocol) is introducing a shift in how developers structure tool access and orchestration for Agents.

• ​Learn more here →​

​2) 4 stages of training LLMs from scratch​

• This visual covers the 4 stages of building LLMs from scratch that are used to make them applicable for real-world use cases.

• These are:

  • Pre-training

  • Instruction fine-tuning

  • Preference fine-tuning

  • Reasoning fine-tuning

• ​Learn more here →​

​3) 3 prompting techniques for reasoning in LLMs​

• A large part of what makes LLM apps so powerful isn't just their ability to predict the next token accurately, but their ability to reason through it.

• This visual covers three popular prompting techniques that help LLMs think more clearly before they answer.

• ​Learn more here →​

​4) Train LLMs using other LLMs​

• LLMs don't just learn from raw text; they also learn from each other:

• Llama 4 Scout and Maverick were trained using Llama 4 Behemoth.

• Gemma 2 and 3 were trained using Google's proprietary Gemini.

• Distillation helps us do so, and the visual below depicts three popular techniques.

• ​Learn more here →​

​5) Supervised & Reinforcement fine-tuning in LLMs​

• RFT lets us transform any open-source LLM into a reasoning powerhouse without any labeled data.

• This visual covers the differences between supervised fine-tuning and reinforcement fine-tuning.

• ​Learn more here →​

​6) Transformer vs. Mixture of Experts

• Mixture of Experts (MoE) is a popular architecture that uses different "experts" to improve Transformer models.

• Experts are like which are feed-forward networks but smaller compared to those in traditional Transformer models.

• ​Learn more here →​

​7) RAG vs Agentic RAG​

• Naive RAG retrieves once and generates once, it cannot dynamically search for more info, and it cannot reason through complex queries.

• Also, there's little adaptability. The LLM can't modify its strategy based on the problem at hand.

• Agentic RAG solves this.

• ​Learn more here →​

​8) 5 Agentic AI design patterns​

• Agentic behaviors allow LLMs to refine their output by incorporating self-evaluation, planning, and collaboration!

• This visual depicts the 5 most popular design patterns employed in building AI agents.

• ​Learn more here →​

​9) 5 levels of Agentic AI systems​

• Agentic systems don't just generate text; they make decisions, call functions, and even run autonomous workflows.

• The visual explains 5 levels of AI agency—from simple responders to fully autonomous agents.

• ​Learn more here →​

​10) Traditional RAG vs HyDE​​

• One critical problem with the traditional ​RAG system​ is that questions are not semantically similar to their answers. As a result, several irrelevant chunks get retrieved during the retrieval due to a higher cosine similarity than the documents actually containing the answer.

• HyDE solves this by generating a hypothetical response first.

• ​Learn more here →​

​11) RAG vs Graph RAG​

• Answering questions that need global context is difficult with traditional RAG since it only retrieves the top-k relevant chunks.

• Graph RAG makes it more robust with graph structures, which helps it build long-range dependencies instead of local text grouping that happens in RAG.

• ​Learn more here →​

​12) KV caching​

• KV caching is a technique used to speed up LLM inference.

• In a gist, instead of redundantly computing KV vectors of all context tokens, we cache them. This saves time during inference.

• ​Learn more here →​

Thanks for reading!