Deep Dive into Solon AI v4.0.3: The Seven‑Layer Java Full‑Stack Agent Framework and How to Get Started

1. Solon AI overview

Solon AI is a Java AI full‑scene development framework built on the Solon application framework. It provides a stack from model integration, knowledge‑base construction, tool invocation to multi‑agent collaboration. Core design: “one code, cross‑model execution”. A dialect‑adaptation layer abstracts model differences, allowing the same code to switch among OpenAI GPT‑4, DeepSeek V3, QWen, etc. Supported model ecosystem includes OpenAI, Claude, Gemini, DeepSeek, QWen, Ollama, Doubao.

2. Four‑layer architecture

Access layer : unified interfaces for Chat, Generate, and Embedding models.

Perception layer : full‑chain RAG support, multi‑format document loading (PDF, Word, Markdown), vector store adapters (Chroma, Milvus, Elasticsearch, Redis, Qdrant), and re‑ranking model integration.

Execution layer : implements MCP, ACP, Function Call and Skills packages; v4.0.3 adds solon‑ai‑loop plugin for loop control in complex workflows.

Brain layer : ReAct self‑reflection, multi‑agent (Team) collaboration, and Harness task orchestration.

3. Changes in v4.0.3

Enhanced MCP compatibility: McpSchema.Sampling annotated with @JsonIgnoreProperties(ignoreUnknown = true) to ignore unknown JSON fields.

Code‑generation capability extracted to independent solon‑ai‑talent‑code plugin (renamed from GenerateTool to Talent), enabling orthogonal enable/disable.

Dialect cache control added in solon‑ai‑core, reducing repeated parsing of model configurations.

Agent tracing improved: ReActTrace and SimpleTrace now include an agentName field.

4. Performance benchmarks

Concurrent throughput up to 700 % higher than a comparable Spring solution (TechEmpower benchmark).

Memory consumption reduced by 50 %.

Package size reduced by 90 %.

Startup time ten times faster.

Benchmarks were run on the TechEmpower third‑party testing platform under realistic conditions.

Solon AI runs on Java 8 through Java 26 and supports GraalVM Native Image.

5. Practical onboarding steps for Java developers

Run a ChatModel demo. The API supports synchronous, streaming, and multimodal (image) calls with minimal code.

Integrate Function Calling to enable agents to invoke local tools such as database queries, internal APIs, or scripts.

Build a RAG knowledge base: ingest documents, generate embeddings, perform vector retrieval, apply re‑ranking, and enhance prompts in a single pipeline.

After completing these steps, developers can extend to multi‑agent collaboration and Harness workflow orchestration.