AI Agents & Agentic Workflows with Spring AI, MCP and Java

Explore how Ollama enables local model execution on your machine with a docker-like workflow. Install, download models, and run via commands, while noting hardware and API considerations.

Create a Gemini API key in Google AI studio by creating a project and copying the key. Explore free-tier limits, billing, and quota options, including models like 3.1 flashlight.

Learn to start simple by interacting with large language models, sending simple prompts through a chat client, understanding the integration, and building concepts before applying to real applications.

Learn how chat options, especially maximum tokens, govern model responses to limit length and cost, with notes on API changes between older and final Spring AI releases.

Explore how temperature controls randomness in model responses, with a 0–2 range; low temperatures produce stable, deterministic outputs for code, while higher values boost creativity for brainstorming.

Configure a Spring app to use multi-model providers by creating separate chat clients for OpenAI, OLAMA, and Google Gen AI, with on-premise data handling and provider-specific profiles.

Explore structured output to convert model responses into Java objects, using native structured output and an advisor to enable reliable JSON formatting in Spring applications.

Learn to use dynamic prompt templates with placeholders to create flexible, reusable prompts for chat clients. Parameterize issue and tone, load templates as resources, and apply system and user prompts.

Explore how prompts evolve into message roles for multi-turn AI chats, define user, assistant, and system roles, and explain including prior messages in the prompt object to preserve context.

Discover how Spring AI's chat memory manages conversation history on the application side, using a conversation id to include prior messages in each request.

Explore a candidate service with a one-to-many candidate and work experience schema, exposing a minimal api that returns candidate data by id via a dto mapper.

Implement a FindJobs API that fetches candidate and job data, scores matches with an LLM, and provides justified reasoning via a career advisor client.

Implement a find jobs api by injecting job, candidate, and advisor clients, retrieving candidate skills, searching jobs by skills, evaluating with the advisor, and returning a sorted match score list.

Create a compare jobs API that fetches candidate and job details, uses an LLM to compare based on skills, location, and salary, and returns a best job recommendation.

Implement the job application submission API, persist the application via the service and DTO mapper, and trigger an asynchronous event to review with an LLM, updating match score and reasoning.

Build and configure the hiring advisor client using a chat client builder, Jackson JSON mapper, and resource loader, loading prompt templates and enabling native structured output for requests.

Enhance an existing microservices app by layering AI on top of candidate, job, and hiring services, providing guidance for candidates and insights for recruiters.

Explore how an MCP server exposes capabilities by grouping features into resources, tools, and prompts. Resources are read-only data, tools perform actions, and prompts are reusable templates.

Clarify why multiple Spring AI modules are included for MCP architecture, using the Spring AI starter MCP server and MCP client modules, with streamable HTTP and LLM integration.

Explore a Spring Boot Maven playground for the MCP, wiring MCP server and host with external LLMs like Google Gen AI or OpenAI, and run independent sectioned hosts and servers.

Explore how the system message defines the LLM's tone and behavior, instructing the use of available tools and proper responses when tools are unavailable, and encouraging experimentation and rule refinement.

MCP is a relatively new specification evolving through community feedback and real world usage, with tools as the foundation for building AI agents; resources and prompts are also important.

Move from data providers to goal-driven, more autonomous systems where the LLM decides what to do and the MCP host executes tool calls via MCP servers.

Explore file operations with the file tools, including creating, reading, deleting, converting csv to json, merging files, and managing runtime directories under target/classes.

Learn how to structure tool parameters using a Java record to pass a single structured input, describe each field, and keep code clean while scaling parameter handling.

Demonstrates structured input workflows using a local server and browser UI to list and schedule reminders with an LLM tool, and verifies timings via server logs.

Explore tool design approaches for an order service, implementing a seven-day cancellation rule and comparing low-level tools, logic-in-tools, and robust failure handling to improve reliability.

Capture tool errors as part of the tool response, avoid exceptions, and use clear messages with status codes such as 400, 401, 404 in MCP-hosted LLM interactions.

Explore pattern 3: design encapsulated and robust tools with try-catch handling, eligibility checks for cancel orders, and user-friendly messages for not found errors, then deploy and test on localhost:8080.

Configure a MCP meta driven project setup that enforces premium user permissions for listing and canceling orders, using a shared dto package, user context, and host-server communication.

Expose real-time progress updates to the UI using a Sync object and Flux, broadcasting via multicast with backpressure in reactive programming.

Create a unique progress token to route per-user updates from the MCP host to the UI, with a flux API filtering by token.