RGA: Beyond RAG with Iterative Retrieval-Generation

prj_rga Project

prj_rga is a project that experiments with a Retrieval-Generation-Augmentation pipeline, going beyond standard RAG. It ran through April 2025 across 5 commits.

RAG vs RGA

graph TD
    subgraph RAG["기존 RAG"]
        A1[Retrieve] --> B1[Augment] --> C1[Generate]
    end

    subgraph RGA["RGA"]
        A2[Retrieve] --> B2[Generate 초안]
        B2 --> C2[Augment 보강]
        C2 --> D2{품질 검증}
        D2 -->|부족| A2
        D2 -->|충분| E2[최종 출력]
    end

Core Idea

1. Retrieve: Search for relevant documents
2. Generate: Produce a draft from the retrieved results
3. Augment: Fill gaps in the draft with additional retrieval
4. Iterate: Repeat until quality is sufficient

Implementation

def rga_pipeline(query: str, max_iterations: int = 3):
    for i in range(max_iterations):
        # Retrieve
        docs = retriever.get_relevant_documents(query)

        # Generate
        draft = llm.invoke(
            f"문서: {docs}\n질문: {query}\n초안을 작성하세요."
        )

        # Augment - 부족한 부분 파악
        gaps = llm.invoke(
            f"초안: {draft}\n부족한 정보를 파악하세요."
        )

        if "충분함" in gaps:
            return draft

        # 부족한 부분으로 추가 검색
        query = gaps
    return draft

A small experiment spanning just 5 commits, but it validated the potential of an iterative retrieve-generate loop.