Building AI Applications ======================== **Why Smarter.** An AI application is more than a model: it is a model plus instructions, knowledge, tools, guardrails and, increasingly, other models. Most courses can show these parts only one at a time, in a different framework each, and only to students who can program. In Smarter, each part is a resource declared in a short YAML manifest, with the same ``apply``, ``describe`` and ``delete`` lifecycle: :doc:`Vectorstores ` on Qdrant or Pinecone for retrieval, :doc:`SQL ` and :doc:`API ` plugins over a :doc:`Connection ` for structured data, :doc:`MCP clients ` for the tools of any MCP server, :doc:`Guardrails ` for safety, and an :doc:`Orchestrator ` for multi-agent workflows. A course can cover the whole landscape of AI application architecture in a term, without first teaching a programming language. Students learn from complete, working applications, not diagrams. Every installation ships with `nineteen built-in LLMClients `__ to take apart, such as `research_analyst `__, which searches the web through the Exa MCP server, reads pages and PDFs, and checks its citations with an output guardrail; `twenty-five ready-made MCP clients `__ for services such as GitHub, Notion, Linear and Stripe; twenty `sample plugins `__; and `Stackademy `__, a fictional university whose course catalog is served both from a database and from a REST API. Smarter also makes the design decisions behind safe AI applications visible, where students can see them and argue about them: - A tool is used only after two decisions: Smarter :doc:`selects a plugin ` when the prompt matches its search terms, and only then does the model decide whether to call it. - A SQL Plugin runs a fixed, `parameterized query `__ that the model can only fill in, so the model never writes SQL that runs. The ``data_analyst`` LLMClient shows the other pattern: a skill that teaches the model a schema, so that it writes SQL for a person to review. - Credentials are :doc:`Secrets `, named in a manifest but never shown to the model, and MCP connections are `restricted `__ to public https endpoints, with an allow-list of the tools the model may use. - An Orchestrator coordinates several LLMClients with one of six `strategies `__: sequential, parallel, supervisor, router, voting and debate. - Every request is :doc:`charged ` with its tokens, so students can compare one assistant with a workflow of several by cost and latency, as well as quality. **What the course covers.** The anatomy of an AI application, and how each part appears as a Smarter resource; models and prompts for applications, including structured output; retrieval: embeddings, chunking, top-k, citations and measuring retrieval quality; structured data, and the safe patterns for letting a model answer from a database; tools, function calling and plugin selection; the Model Context Protocol, and trusting an MCP server; tool safety and prompt injection through retrieved content; orchestration, from workflows to agents; evaluation with test sets, rubrics and LLM judges; and the production concerns of guardrails, budgets, cost accounting, the journal and deployment. **Look at.** - `llmclient/data/llm-clients/ `__, especially ``customer_support``, ``research_analyst`` and ``data_analyst``. - `mcpclient/data/mcpclients/ `__ and `plugin/data/ `__. - `vectorstore/data/vectorstores/ `__, one manifest for each kind of vector database. Each course plan lists every manifest and file that it uses under *Smarter in the Source Code*. **Read.** :doc:`/smarter-resources/smarter-vectorstore`, :doc:`/smarter-resources/smarter-vectorsearch`, :doc:`/smarter-resources/smarter-mcpclient`, :doc:`/smarter-resources/smarter-connection`, :doc:`/smarter-resources/plugin/plugin/sql`, :doc:`/smarter-resources/smarter-orchestrator`, :doc:`/smarter-resources/plugin/nlp`. **Exercises.** - Take the ``research_analyst`` manifest apart: list every resource it uses, what each does, and draw how a question flows through it. Confirm each step in the workbench's Console. - Load a course syllabus and handbook into a Vectorstore, and build a teaching assistant that answers questions about them, with citations. - Answer business questions in plain English over the Stackademy database, with the ``stackademy_sql`` SQL Plugin, and compare it with the ``data_analyst`` LLMClient, which writes SQL for a person to review. Which would you let a manager rely on, and why? - Build the same workflow as a single LLMClient and as an Orchestrator of several, and compare cost, latency and quality. **Course plans.** Three complete, ready-to-teach plans, each with a schedule, graded assignments and rubrics, two capstone projects, recommended texts, further study and technical prerequisites. See :doc:`course-formats` for what the formats have in common. .. toctree:: :maxdepth: 1 Classroom bootcamp: 8 hours Two-week online course: 6 hours Four-week online course: 12 hours See the :doc:`Smarter Instructor Guide ` to install Smarter and set up a class.