AI System Architecture Components (2026): RAG, MCP & Agents

What Are the Components of an AI System Architecture?

18 entries

A modern AI application is more than a language model generating a response. It needs systems around the model to retrieve information, understand context, use tools, take actions and stay within boundaries. A LinkedIn post by Keith R. Worfolk (27 October 2026) maps nine concepts that he says are pieces of one architecture, not separate trends: embedding models, vector databases, retrieval-augmented generation (RAG), prompt engineering, semantic caching, the Model Context Protocol (MCP), AI agents, function calling and guardrails. He sketches the flow as input, context, retrieval, reasoning, tool use, action, validation and output, and argues that the model is one component while the architecture is the product. This page follows those nine components, lists primary references and tools for each, and links related pages on this site. Descriptions are paraphrased from each source, are not rankings, and no performance claims are verified here.

01RAG, vector databases, MCP, agents, guardrails: the same architecture (LinkedIn post by Keith R. Worfolk)The post that maps the nine components and the input-to-output flow. 02AI ArchitectureData, models, RAG, agents and gateways, on this site. 03AI System Design Implementation TechniquesTen system design building blocks for AI systems, on this site. 04Enterprise AI System Design ConceptsTwelve distributed-systems concepts for production AI, on this site. 05Why Enterprise AI Needs a Connected ArchitectureMCP, gateways and A2A together, on this site. 06Embedding ModelsText embedding models, on this site. 07Vector DatabasesVector database directory, on this site. 08Production-Ready RAG ArchitectureRAG pipeline design and checklist, on this site. 09Open-Source RAG StackSeven layers of an open-source RAG system, on this site. 10MCP Design PatternsPatterns for MCP servers, on this site. 11Top MCP ServersMCP servers directory, on this site. 12AI Agent FrameworksFrameworks for building agents, on this site. 13Agent MemoryMemory layers for agents, on this site. 14Architectural Patterns for Agentic AISix patterns for structuring agents, on this site. 15Guardrails for Agentic AIGuardrails for autonomous agents, on this site. 16AI GatewaysRouting, caching and cost governance, on this site. 17Token Optimization ToolsCompression, caching and routing to cut tokens, on this site. 18AI Governance FrameworksRegulation, standards and controls, on this site.

The Nine Components

9 entries

The post's nine components in its order, each with a paraphrased role and a reference to start from.

191. Embedding modelsTurn data into meaningful vectors. See the MTEB leaderboard, which lets you browse benchmark results for many embedding models. 202. Vector databasesStore vectors for semantic search; options range from pgvector in Postgres to dedicated engines. 213. RAGRetrieves relevant information before generation. The original RAG paper is Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. 224. Prompt engineeringShapes instructions and context. OpenAI's guide covers strategies and tactics for better results with language models. 235. Semantic cacheReduces repeated work by reusing answers to similar queries; Redis describes LangCache as semantic caching for LLM applications. 246. MCPConnects AI applications to tools and external systems through an open protocol. 257. AI agentsAdd planning, memory, reasoning and autonomy. Anthropic's guide covers building effective agents. 268. Function callingLets models trigger actions and APIs; OpenAI describes connecting models to external data and systems. 279. GuardrailsValidate inputs, outputs and behavior; NVIDIA's NeMo Guardrails describes programmable guardrails for LLM-based systems.

Embeddings, Vector Search and Retrieval

8 entries

Building blocks for the retrieval path.

28Sentence TransformersPython library for using and training embedding and reranking models. 29FAISSLibrary for similarity search and clustering of dense vectors. 30MilvusOpen-source vector database for scalable vector search. 31QdrantQdrant describes an open-source vector search engine written in Rust. 32WeaviateWeaviate describes an AI database for AI-native applications. 33ChromaChroma describes open-source search infrastructure for AI. 34LlamaIndexFramework for building RAG and agent applications over your data. 35Haystackdeepset describes an open-source AI framework with modular building blocks for production applications.

Prompting and Caching

9 entries

Instruction design and ways to avoid repeated work.

36Prompt Engineering GuideA comprehensive overview of prompt engineering. 37Claude prompt engineering overviewAnthropic documentation on when prompt engineering is the right solution and on Claude prompting techniques. 38Gemini prompt design strategiesGoogle documentation on prompt design for the Gemini API. 39DSPyDSPy describes a framework for programming, rather than prompting, language models. 40GPTCacheSemantic cache for LLMs; GitHub lists 0.1.44 (August 2024) as the latest release, so check activity before adopting. 41Redis LangCacheRedis describes semantic caching that skips repeated model calls. 42LiteLLM cachingLiteLLM documentation on caching responses in its proxy. 43OpenAI prompt cachingOpenAI describes how prompt caching reduces latency and cost for long prompts. 44Claude prompt cachingAnthropic describes caching prompt prefixes to cut costs and latency.

Tools, MCP, Agents and Function Calling

7 entries

How models reach tools and take actions.

45MCP specificationThe Model Context Protocol specification. 46MCP specification repositorySpecification and documentation for the Model Context Protocol on GitHub. 47Tool use with ClaudeAnthropic documentation on connecting Claude to external tools and APIs. 48Function calling with the Gemini APIGoogle documentation for function calling with Gemini. 49ReAct: Synergizing Reasoning and Acting in Language Models (arXiv)Yao et al. on interleaving reasoning and actions in language models. 50Toolformer (arXiv)Schick et al. on language models teaching themselves to use tools. 51LangGraphLangChain describes designing agents that reliably handle complex tasks; GitHub lists 1.2.19 (30 October 2026).

Guardrails, Validation and Observability

9 entries

Checks on inputs and outputs, and measurement of behavior.

52NeMo GuardrailsToolkit for adding programmable guardrails to LLM-based conversational systems; GitHub lists v0.24.1 (16 September 2026). 53Guardrails AILibrary for adding guardrails to large language models; GitHub lists v0.11.0 (14 August 2026). 54Llama Guard (arXiv)Paper on an LLM-based input-output safeguard for human-AI conversations. 55Purple LlamaMeta's set of tools to assess and improve LLM security. 56OWASP Top 10 for LLM ApplicationsOWASP list of top risks for large language model applications. 57RagasRagas describes an evaluation framework for AI applications. 58LangfuseLangfuse describes an open platform to trace, evaluate and improve AI agents. 59OpenTelemetryVendor-neutral observability framework for traces, metrics and logs. 60NIST AI Risk Management FrameworkNIST framework for managing AI risk.

Frequently Asked Questions

What are the components of an AI system architecture?

The post this page follows names nine: embedding models, vector databases, RAG, prompt engineering, semantic caching, MCP, AI agents, function calling and guardrails, working together around the model.

How do the components fit together?

The post sketches a flow of input, context, retrieval, reasoning, tool use, action, validation and output: embeddings and vector databases support retrieval, prompts shape context, MCP and function calling connect tools, agents plan and act, and guardrails validate inputs, outputs and behavior.

What is the difference between function calling and MCP?

Function calling lets a model request that an application run a function or API, while MCP is an open protocol for connecting AI applications to tools and data sources in a standard way. They are often used together. See the MCP Design Patterns page.

What does a semantic cache do?

It reuses answers to semantically similar queries, which cuts repeated model calls, latency and cost, with the risk of returning a stale or mismatched answer. See the Token Optimization Tools and AI Gateways pages.

Why do agents need guardrails?

Agents can plan and take actions through tools, so inputs, outputs and behavior should be validated and permissions limited. See the Guardrails for Agentic AI and AI Governance Frameworks pages.

Which component is hardest to get right in production?

The post asks readers this question without a single answer. In practice retrieval quality, evaluation and guardrails are common trouble spots; see the Production-Ready RAG Architecture page.

Does this page rank tools?

No. Tools are grouped by role, descriptions follow each source's own pages, and no performance claims are verified here.

How current is this page?

Links and release facts were checked on 1 November 2026 from GitHub. GPTCache's latest release is from August 2024, as its card notes, and LinkedIn may ask visitors to sign in. Descriptions follow each source's own pages and are not rankings.