Prompt Repetition

langchain4j-community-prompt-repetition is an optional community module that provides ready-made prompt repetition integrations for two LangChain4j integration points:

• AI Services input guardrails
• RAG query transformation

It is inspired by the paper Prompt Repetition Improves Non-Reasoning LLMs, which reports gains on a range of non-reasoning workloads. In LangChain4j, the module exposes the core repeated-input transformation through framework-native components and adds conservative defaults for real applications.

The module is experimental, and its effectiveness is workload-dependent. Validate it on your own prompts, models, and tasks before broad rollout.

What It Is

Prompt repetition rewrites text in the following form:

Q -> Q\nQ

In LangChain4j, this can be applied in two different places:

• Before a non-RAG AI Services call, using PromptRepeatingInputGuardrail
• Before retrieval in an advanced RAG pipeline, using RepeatingQueryTransformer

For RAG, the repetition should be applied only to the retrieval query, not to the final augmented prompt sent to the model.

Maven Dependencies

If you already use community modules, importing the community BOM is recommended:

<dependencyManagement>
    <dependencies>
        <dependency>
            <groupId>dev.langchain4j</groupId>
            <artifactId>langchain4j-community-bom</artifactId>
            <version>${latest version here}</version>
            <type>pom</type>
            <scope>import</scope>
        </dependency>
    </dependencies>
</dependencyManagement>

Then add the prompt repetition module:

<dependency>
    <groupId>dev.langchain4j</groupId>
    <artifactId>langchain4j-community-prompt-repetition</artifactId>
</dependency>

You can also declare the module directly:

<dependency>
    <groupId>dev.langchain4j</groupId>
    <artifactId>langchain4j-community-prompt-repetition</artifactId>
    <version>${latest version here}</version>
</dependency>

Components

The module provides three main APIs:

• PromptRepeatingInputGuardrail repeats eligible single-text user input before the model is called
• RepeatingQueryTransformer repeats the retrieval query in advanced RAG pipelines
• PromptRepetitionPolicy contains the shared repetition rules used by both integrations

All of these APIs are marked as @Experimental.

Non-RAG Usage

For non-RAG AI Services calls, attach PromptRepeatingInputGuardrail to the AiServices builder:

PromptRepetitionPolicy policy = PromptRepetitionPolicy.builder()
        .mode(PromptRepetitionMode.AUTO)
        .maxChars(8_000)
        .build();

Assistant assistant = AiServices.builder(Assistant.class)
        .chatModel(chatModel)
        .inputGuardrails(new PromptRepeatingInputGuardrail(policy))
        .build();

This is the preferred integration point when you want to rewrite the user input before the model call and you are not operating on an augmented RAG prompt.

RAG Usage

For RAG, repeat the retrieval query only:

PromptRepetitionPolicy policy = PromptRepetitionPolicy.builder()
        .mode(PromptRepetitionMode.AUTO)
        .maxChars(8_000)
        .build();

RetrievalAugmentor retrievalAugmentor = DefaultRetrievalAugmentor.builder()
        .queryTransformer(new RepeatingQueryTransformer(policy))
        .build();

This keeps the transformation inside the retrieval stage and avoids duplicating the final prompt after retrieved content has already been injected.

Modes

PromptRepetitionPolicy supports three modes:

• NEVER: disables repetition
• ALWAYS: repeats eligible input
• AUTO: conservative mode that skips already repeated text, very long input, and prompts that appear to request explicit reasoning

AUTO is the safest starting point for evaluation.

Safety and Constraints

• PromptRepeatingInputGuardrail only rewrites eligible single-text user input
• It is not intended to be the main integration point for multimodal requests
• By default, the guardrail skips requests when RAG augmentation has already happened
• In RAG setups, use RepeatingQueryTransformer to repeat the retrieval query instead of repeating the final augmented prompt
• The module is experimental, so APIs and behavior may change in future versions

When to Use It

Use this module when you want a ready-made way to apply prompt repetition in LangChain4j, not when you want a universal default prompt policy.

• Start with PromptRepetitionMode.AUTO
• Prefer it for non-reasoning or low-reasoning workloads first
• Evaluate it with A/B tests on your own prompts, models, and tasks
• Keep the default safety constraints unless you have a clear reason to relax them
• Treat improvements as workload-dependent rather than guaranteed