Agent, RAG, MCP & ML

Regenerate Background

Anatomy of an AI agent

How a modern LLM agent is put together: sense inputs, think (deterministic, LLM, or hybrid), act and observe in a loop, finish conditions, plus optional evaluate, memory, description, planning, chain-of-thought, ask — and how RAG grounds answers.

RAG Design Patterns (Chunking, Retrieval, Ranking)

Twelve practical retrieval-augmented generation patterns: from fixed-size and context-aware chunking and contextual retrieval to re-ranking, query expansion, multi-query, agentic and self-reflective RAG, graphs, hierarchy, late chunking, and domain-tuned embeddings — plus three stacks teams actually ship and mistakes to avoid.

Anatomy of RAG

Typical vector RAG pipeline: ingest, chunk, embed, store, query, retrieve, rerank, augment, generate — plus core components (data layer, chunking, embeddings, storage, retriever, generator).

LLM Engines

Reference list of LLM and embedding models by provider: OpenAI, Google, Cohere, Voyage AI, Qwen, DeepSeek, Kimi. Generative models, reasoners, embeddings, rerankers.

Agent loop patterns

Nine ReAct-style agent loop patterns: think–act–observe and extensions — dialogue, description, multi-tool, reflection, memory, planning, chain-of-thought, and learning — with pros, cons, and model notes.

Production Agent-RAG Architectures

Enterprise knowledge copilot as agentic RAG: ingest and hybrid retrieval, optional ReAct, grounding and citations — workflow, typical tools, and when it fits.

Model Context Protocol (Anthropic)

The Model Context Protocol (MCP) as defined by Anthropic is a standardized client–server protocol and runtime ecosystem for connecting AI applications to external tools, data sources, and prompts. It is not an architectural pattern for designing agents; it is the concrete protocol and specification that enables applications like Claude, Cursor, and other MCP clients to discover and use Resources, Tools, and Prompts exposed by MCP servers.