Skip to content

README.md

Latest commit

 

History

History
72 lines (52 loc) · 1.76 KB

README.md

File metadata and controls

72 lines (52 loc) · 1.76 KB
Raw

LangChain4j Community Prompt Repetition

Prompt repetition

This module includes optional prompt repetition components.

Prompt repetition rewrites:

Q -> Q\nQ

Use it only where it helps your workload.

Components

  • PromptRepeatingInputGuardrail for non-RAG user input.
  • RepeatingQueryTransformer for RAG retrieval queries.
  • PromptRepetitionPolicy shared by both.

Modes

  • NEVER: disable.
  • ALWAYS: repeat when eligible.
  • AUTO: conservative mode:
    • skip already repeated input
    • skip very long input (maxChars)
    • skip reasoning-intent prompts (reasoningKeywords)

Non-RAG usage

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

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

RAG usage

Repeat the retrieval query only. Do not repeat the full augmented prompt.

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

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

Safety behavior

  • Input guardrail rewrites only single-text user messages.
  • Input guardrail skips by default when RAG augmentation is detected.
  • Idempotence is built in:
    • first pass: Q -> Q\nQ
    • repeated pass: SKIPPED_ALREADY_REPEATED

Rollout checklist

  1. Start with AUTO.
  2. Keep maxChars conservative.
  3. Tune reasoningKeywords for your workload.
  4. Track apply/skip reasons during A/B evaluation.