Using PromptTemplate in Spring AI for Variable Injection

Direct Answer

PromptTemplate in Spring AI takes a Mustache-style string with {variable} placeholders and a Map<String, Object> of values, producing a typed Prompt you pass to ChatModel or ChatClient. SystemPromptTemplate does the same for system-role messages. ChatClient.builder().defaultSystem(...) sets a fixed system instruction for every call. This approach replaces ad-hoc string concatenation with versioned, testable prompt objects.

Deep Dive & Architecture

Zero-shot: no examples in the prompt — just the instruction.

Fast and cheap; works for simple tasks the model has seen in training. Few-shot: include 2–5 worked examples of input→output pairs.

Dramatically improves format consistency and domain accuracy without fine-tuning. Chain-of-thought (CoT): instruct the model to reason step by step before giving the final answer. Use for multi-step logic, math, or any task where correctness depends on intermediate reasoning. Cost increases with output tokens. For Java teams: zero-shot for simple extraction, few-shot for consistent JSON output, CoT for complex domain reasoning.

Treating prompts as static strings hardcoded in Java — missing variable injection, system vs user role separation, and testability

Common Mistakes & Pitfalls

Building prompts with String.format() or concatenation — no structure, hard to test, prone to injection if user input isn't sanitized