Skip to main content

Orchestration Family (G3)

These elements coordinate multiple components into working systems.

You can't orchestrate one thing—this family only exists through combination. Context windows hold the information, RAG fills them intelligently, frameworks provide the plumbing, and protocols standardize connections.

ElementNameRowDescription
CwContext WindowsPrimitivesThe limited space an LLM can "see"
RgRAGCompositionsRetrieval-Augmented Generation
FwFrameworksDeploymentPlatforms connecting AI components
McMCPEmergingModel Context Protocol and standards

Cw — Context Windows

Position in Periodic Table:

G3: Orchestration Family
┌──────────────────────────┐
│ → [Context Windows] │ Row 1: Primitives
│ RAG │ Row 2: Compositions
│ Frameworks │ Row 3: Deployment
│ MCP │ Row 4: Emerging
└──────────────────────────┘

What It Is

Context windows are the limited space an LLM can "see" at once. Managing what goes into context is fundamental orchestration. Every token costs money and attention. Think of it as the model's working memory.

Why It Matters

Understanding context windows is crucial because:

  • There's a hard limit on what the model can process
  • Not all context is equal—position and relevance matter
  • Token costs scale linearly with context size
  • "Lost in the middle" phenomenon affects retrieval
  • Context management is often the difference between success and failure

Context Window Sizes (2026)

ModelContext Window
GPT-4 Turbo128K tokens
Claude 3200K tokens
Gemini 1.51M+ tokens
Llama 38K-128K tokens

What Consumes Context

ComponentDescription
System promptInstructions, persona, rules
Conversation historyPrevious messages
Retrieved contextRAG results, documents
User inputCurrent message
OutputGenerated response

The "Lost in the Middle" Problem

Research shows models pay most attention to:

  1. The beginning of context (primacy)
  2. The end of context (recency)
  3. Less attention to the middle

Implication: Put important information at the start or end, not buried in the middle.

Context Management Strategies

  1. Compression: Summarize long content to preserve meaning in fewer tokens
  2. Selection: Only include what's relevant to the current query
  3. Chunking: Break documents into pieces, retrieve only relevant chunks
  4. Sliding Window: Keep recent history, summarize or drop older content
  5. Hierarchical: Summary of full context + detailed recent context

Tier Relevance

TierExpectation
FoundationUnderstand context limits and token costs
PractitionerImplement context management strategies
ExpertOptimize context for cost/quality tradeoffs

Rg — RAG

Position in Periodic Table:

G3: Orchestration Family
┌──────────────────────────┐
│ Context Windows │ Row 1: Primitives
│ → [RAG] │ Row 2: Compositions
│ Frameworks │ Row 3: Deployment
│ MCP │ Row 4: Emerging
└──────────────────────────┘

What It Is

RAG (Retrieval-Augmented Generation) combines retrieval with generation. A question comes in, relevant context is retrieved (via embeddings/vector DB), the prompt is augmented with that context, and the LLM generates a grounded answer.

Why It Matters

RAG solves critical LLM limitations:

  • Hallucination: Grounds answers in retrieved facts
  • Stale knowledge: Access up-to-date information
  • Domain specificity: Query your proprietary data
  • Traceability: Know where answers came from

The RAG Pipeline

User Query → Embed → Search Vector DB → Retrieve Docs

Retrieved Docs + Query → Augmented Prompt → LLM → Grounded Response

Key Components

ComponentPurpose
Embedding modelConvert query to vector
Vector databaseStore and search documents
RetrieverFind relevant documents
RerankerImprove retrieval ordering
GeneratorLLM that produces the answer

Chunking Strategies

StrategyDescriptionBest For
Fixed sizeSplit every N charactersSimple, predictable
SentenceSplit on sentence boundariesReadability
ParagraphSplit on paragraph breaksStructured documents
SemanticSplit on topic changesMixed content
RecursiveTry multiple strategiesGeneral purpose

Advanced Techniques

TechniqueDescription
Hybrid searchCombine vector + keyword search
Query expansionRephrase query for better retrieval
RerankingUse cross-encoder to reorder results
Multi-queryGenerate multiple queries, merge results
Self-RAGModel decides when to retrieve

Common Pitfalls

  • Poor chunking: Chunks that split important context
  • No reranking: First result isn't always best
  • Ignoring metadata: Missing filtering opportunities
  • Too few/many chunks: Balance context vs. noise
  • No evaluation: Not measuring retrieval quality

Tier Relevance

TierExpectation
FoundationUnderstand the RAG pattern conceptually
PractitionerBuild complete RAG pipelines
ExpertOptimize retrieval quality and implement advanced patterns

Fw — Frameworks

Position in Periodic Table:

G3: Orchestration Family
┌──────────────────────────┐
│ Context Windows │ Row 1: Primitives
│ RAG │ Row 2: Compositions
│ → [Frameworks] │ Row 3: Deployment
│ MCP │ Row 4: Emerging
└──────────────────────────┘

What It Is

AI frameworks are platforms like LangChain, LlamaIndex, and others that provide the plumbing to connect AI components. They handle the complexity of building and deploying AI systems.

Why It Matters

Frameworks accelerate development by:

  • Providing battle-tested implementations
  • Abstracting away boilerplate
  • Enabling rapid prototyping
  • Standardizing patterns across teams
  • Integrating with various providers

Major Frameworks (2026)

FrameworkFocusBest For
LangChainGeneral-purpose chainsFlexible AI applications
LlamaIndexData indexing/retrievalRAG-heavy applications
HaystackNLP pipelinesSearch and QA
Semantic KernelMicrosoft ecosystem.NET/enterprise
LangGraphStateful agentsComplex agent workflows
CrewAIMulti-agentAgent teams

Common Abstractions

AbstractionPurpose
ChainsSequence of operations
AgentsAutonomous decision-making
ToolsFunctions agents can call
MemoryConversation state
RetrieversDocument retrieval

When to Use Frameworks

Use frameworks when:

  • Rapid prototyping is priority
  • Using standard patterns (RAG, agents)
  • Team benefits from shared abstractions
  • Integrating multiple providers

Consider direct APIs when:

  • Simple, single-model use case
  • Maximum control needed
  • Framework overhead is concern
  • Highly custom patterns

Framework Tradeoffs

AspectProsCons
AbstractionFaster developmentLess control
UpdatesCommunity improvementsBreaking changes
DebuggingLogging built-inStack traces can be opaque

Tier Relevance

TierExpectation
FoundationAwareness of major frameworks
PractitionerProficiency in at least one framework
ExpertEvaluate frameworks, know when to use vs. build

Mc — MCP

Position in Periodic Table:

G3: Orchestration Family
┌──────────────────────────┐
│ Context Windows │ Row 1: Primitives
│ RAG │ Row 2: Compositions
│ Frameworks │ Row 3: Deployment
│ → [MCP] │ Row 4: Emerging
└──────────────────────────┘

What It Is

MCP (Model Context Protocol) and similar standards are emerging protocols for how AI systems connect to tools and data sources. Think of it as the USB standard of AI integration.

Why It Matters

Without standards, every AI-tool integration is custom. MCP and similar protocols enable:

  • Reusable tool implementations
  • Interoperability across AI systems
  • Easier ecosystem development
  • Reduced integration complexity

The Problem MCP Solves

Before: Each AI system needs custom code for each tool

Claude ──custom──> Slack
Claude ──custom──> GitHub

GPT ────custom──> Slack (different code!)
GPT ────custom──> GitHub (different code!)

After: Tools expose standard interface, any AI can use

Claude ──MCP──┐
├──> Slack Server (one implementation)
GPT ────MCP──┘

MCP Architecture

ComponentRole
HostAI application (Claude, IDEs, etc.)
ServerTool provider (Slack, GitHub, DB)
ProtocolStandardized communication format
ResourcesData the server exposes
ToolsActions the server enables

Current State (2026)

MCP was introduced by Anthropic in late 2024 and is gaining adoption. The ecosystem is growing, but:

  • Not all providers support it yet
  • Some features are still evolving
  • Alternative standards may emerge
  • Best practices are still forming

Evaluating Protocol Adoption

FactorQuestion
EcosystemAre servers available for your tools?
SupportDoes your AI platform support it?
MaturityIs the protocol stable enough?
SecurityDoes it meet your security requirements?

Tier Relevance

TierExpectation
FoundationAwareness of emerging standards
PractitionerUse existing MCP servers
ExpertEvaluate protocols, build custom servers