The modern AI systems roadmap
An AI-powered system is ordinary software around a probabilistic model. The model predicts useful output; your application owns data, tools, permissions, state, validation, cost, latency, evidence, and recovery.
Current baseline, August 2026
| Layer | Current production baseline | Important note |
|---|---|---|
| Spring AI | 2.0.0 GA | Spring Boot 4.0/4.1 and Spring Framework 7 |
| Java | 21+ | Use the version supported by your Boot line |
| OpenAI | Responses API; GPT-5.6 family | Route Sol/Terra/Luna by workload role |
| Gemini | Gemini 3.6 Flash; 3.5 Flash-Lite | Stable model IDs can still deprecate later |
| Ollama | 0.32.5 | Local runtime; capabilities depend on the chosen model |
| Google embeddings | gemini-embedding-2 |
GA multimodal embeddings, 128–3072 dimensions |
| OpenAI embeddings | text-embedding-3-small/large |
1536/3072 default dimensions |
| Local embeddings | embeddinggemma, qwen3-embedding |
Keep index and query model identical |
| Vector search | pgvector HNSW or managed equivalent | Measure recall, latency, filters, and operations |
| Agent integration | Tool calling and MCP | Tools need authorization, schemas, budgets, and audit |
“Latest” means the newest compatible, stable set that fits quality, latency, cost, privacy, and lifecycle requirements. A flagship model is not the default for every classification or extraction request.
The seven layers
experience → orchestration → model gateway → retrieval/tools → enterprise systems → safety/evaluation → operations
Spring AI sits mainly in orchestration and provider integration. Ollama, OpenAI, and Gemini are inference providers. A vector database is retrieval infrastructure. None of them alone creates a trustworthy product.
Build sequence
- Define the user decision or job.
- Create an evaluation dataset before choosing architecture.
- Establish a simple prompt baseline.
- Add structured output when software consumes the result.
- Add RAG only when external knowledge is needed.
- Add tools only when the system must act or obtain live state.
- Add an agent loop only when a fixed workflow cannot express the task.
- Fine-tune only after measured prompt/RAG/tool limitations.
Feynman check
Explain the system as a librarian, a writer, and a careful operator. Retrieval finds evidence; the model writes; application code checks and acts.