Most of the vocabulary around modern AI describes a small number of jobs. Once those jobs are separated, architecture becomes easier to reason about and less likely to turn into product shopping.
| Job | Concept | Representative implementation |
|---|---|---|
| Generate or transform content | Foundation model, LLM, multimodal model | A hosted model API or an open-weight model |
| Supply relevant evidence | Search, embeddings, retrieval-augmented generation | PostgreSQL with pgvector or a dedicated search service |
| Read or change a system | API, tool, function calling | A narrowly scoped business-system operation |
| Standardize integrations | Model Context Protocol | A compatible host, client and server |
| Choose the next action | Agent orchestration | Custom control logic or a framework such as LangGraph |
| Change model behaviour | Post-training, fine-tuning, distillation | A training pipeline with separate evaluation data |
| Run the model | Inference runtime and serving | llama.cpp, MLX-based tooling or a managed endpoint |
An embedding represents an item numerically so a search system can compare it with other items. It is useful for similarity, not proof of correctness. RAG, or retrieval-augmented generation, supplies selected evidence during a request. Fine-tuning changes learned parameters. A tool exposes a capability. An agent uses model output to help choose what happens next.
Frameworks package some of these mechanisms. LangChain provides integrations and agent abstractions, LangGraph supports orchestration and persistent execution, and LlamaIndex provides data and retrieval-oriented capabilities. Their responsibilities overlap. Start with the work you need removed, then decide whether a framework removes enough of it to justify the behaviour your team must still understand and operate. [13–14]
For repeated extraction, start with a model call and validation. For a stable approval process, start with an explicit workflow. Add retrieval when the evidence is too large, changes often or needs permission-aware selection. Add an agent when model-directed choice among next steps creates measurable value. MCP becomes relevant when several compatible applications need reusable access to the same integrations.
Every added component has an operational cost as well as a learning cost. It brings configuration, failure modes and a maintenance obligation. A direct integration that works may be a better starting point than a generalized platform whose benefit has not yet appeared.