Inject System Prompts Dynamically

Beyond the global system-prompt and agent prompts, erupt-ai offers two finer-grained injection points: the application-level SystemPromptProvider and the request-level LlmRequest.

Dynamic System Prompt Injection

TIP

Supports dynamic extension of the system prompt — injected on demand when a user sends a question, precisely controlling token consumption while improving answer relevance and accuracy.

1. Implement the SystemPromptProvider interface
2. Call registerProvider
3. Implement the getPrompt method

@Component
public class OrderAiPrompt implements SystemPromptProvider {

    @PostConstruct
    public void init() {
        SystemPromptProvider.registerProvider(this);
    }

    @Override
    public String getPrompt() {
        return """
                ## Order Assistant
                When the user asks about order-related questions, prioritize querying data with the queryOrder tool before answering.
                Do not fabricate order information. Amount unit is CNY yuan; time format is yyyy-MM-dd HH:mm.
                """;
    }

}

LlmRequest Request-Level Extensions v2.0.0+

LlmRequest supports injecting additional prompts and behavior switches into a single LLM call — useful for dynamic, context-aware scenarios.

Field	Description
agentPrompt	The agent-role prompt for this call — temporarily overrides the agent's system prompt
contextPrompt	Supplemental context appended to the conversation (e.g., a summary of the currently visible data)
thinking	Whether to enable the model's thinking mode (default false) v2.1.0+
responseFormat	Response format: text (default) or json_object v2.1.0+
tools	Request-scoped tool objects (langchain4j @Tool methods) driving a ReAct loop for this call only — independent from autoCallTool, which exposes the global toolbox instead v2.1.0+

These fields are passed by the frontend or integration layer when calling the chat API. Combined with @Erupt(prompt = "...") and @Edit(prompt = "..."), they allow the AI to understand the business semantics of each entity and field.

Contributors

YuePeng

Claude Opus 5 (1M context)

Changelog

Last edited 2 minutes ago

View full history

f3752b9-filter: add a role-based row-level data filtering example