AI Agent Java Demo (Local LLaMA)

This repository contains a standalone Java example demonstrating direct integration with a local LLaMA-based LLM.

The purpose of this example is to demonstrate the core integration boundary between a Java application and a Large Language Model (LLM):

• programmatic prompt construction
• plain HTTP communication
• explicit request/response handling
• no frameworks or SDK abstractions

This project intentionally focuses on clarity and mechanics rather than features, scalability, or production concerns.

---

Repository Layout

ai-agent-java-demo/
  build.gradle
  settings.gradle

  src/
    main/
      java/
        com/
          celso/
            agentsjavademo/
              ProgrammaticPrompting.java
              LLM.java
              Message.java
     test/
      java/
        com/
          celso/
            agentsjavademo/
              MessageTest.java
              PromptBuilderTest.java

  README.md
  LICENSE
  .gitattributes

---

Getting Started

Requirements

• Java 17+
• A local LLaMA server running on http://localhost:8080 (TinyLlama model is ideal)

Run

This project requires a locally running LLaMA-compatible server exposing an OpenAI-compatible API at:

http://localhost:8080

Step 1 — Start the local LLM server Start a local LLaMA server (for example, using llama.cpp / llama-server) with a TinyLLaMA model.

Example command:

./llama-server
--model /path/to/tinyllama.gguf
--port 8080
--ctx-size 2048
--threads 4

The exact installation path and model location are environment-specific.

Step 2 — Run the Java application Once the LLM server is running, execute:

./gradlew run

and wait a few seconds or minutes (depending on your machine and prompt/system prompt).

Its System Prompt: "You are a cook ready to help with everyday cooking questions." The LLaMA model will answer the prompt: "What are the ingredients of a cake?"

Note: TinyLlama can make mistakes. Check out this important information.

---

Run tests

The tests do not require a running LLM. HTTP-level tests can be added later using an in-memory mock server.

./gradlew test

---

How this works

• The Java application builds the prompt programmatically
• The prompt is sent to a locally running LLaMA server via HTTP
• The response is received and processed by the Java runtime
• There is no browser, notebook, or interactive UI involved

This is an example of LLM integration at the application level, similar to how LLMs are used in production systems.

Notes

This project is inspired by concepts from the "AI Agents in Java" course by Jules White.

The original examples rely on OpenAI API keys. In this repository, the code was reorganized and adapted to work with a locally hosted LLaMA-compatible LLM (TinyLLaMA).

---

Author

This project was developed by an engineer and data scientist with a background in:

• Postgraduate degree in Data Science and Analytics (USP)
• Bachelor's degree in Computer Engineering (UERJ)
• Special interest in statistical models, interpretability, and applied AI

---

Contact

• LinkedIn
• Or open an issue