← 回到學習日誌 Back to Journal
學習札記 / Learning Note

下一代 RAG，不只是更多 ChunksNext-Generation RAG Is More Than More Chunks

大型知識庫搜尋不該只找到看起來相似的段落，而要能沿著文件結構找到真正相關的頁面與證據。Large knowledge bases should retrieve more than similar-looking paragraphs. They should follow document structure to the relevant pages and evidence.

文件章節、樹狀索引與證據連結的三段分鏡
01 保留原始章節結構Preserve document structure02 沿樹狀索引推理Reason through a tree index03 回到頁面與證據Return to pages and evidence
中文版 Chinese version

相似不代表相關

傳統切塊搜尋容易找到語意相近的文字，卻可能失去章節關係、前後條件與完整證據。對大型技術文件來說，結構本身就是資訊。

PageIndex 的做法

PageIndex 保留文件的樹狀章節，讓 LLM 先判斷應該走到哪一個分支，再找出真正相關的頁面。這比先切成大量 chunks 更接近人閱讀長文件的方法。

搭配受控的知識循環

再用 Wiki 累積經過驗證的摘要、概念和交叉連結，搜尋就不只是一次性回答，而能逐步形成可追溯的知識。

English version

Similarity Is Not Relevance

Chunk search can find semantically similar text while losing chapter relationships, conditions, and evidence. In technical documents, structure is information.

How PageIndex Works

PageIndex preserves a chapter tree. The LLM first decides which branch to follow, then retrieves the relevant pages. This resembles how people navigate long documents.

A Controlled Knowledge Loop

A wiki can accumulate reviewed summaries, concepts, and links. Retrieval then becomes traceable knowledge rather than a one-off answer.

PageIndexRAGKnowledge Architecture