Auto-sync
This commit is contained in:
25
wiki/concepts/Source Grounding.md
Normal file
25
wiki/concepts/Source Grounding.md
Normal file
@@ -0,0 +1,25 @@
|
||||
---
|
||||
title: "Source Grounding"
|
||||
type: concept
|
||||
tags: [ai, rag, hallucination-prevention]
|
||||
---
|
||||
|
||||
## Definition
|
||||
Source Grounding(源 grounding)是一种将 AI 回答严格锚定在用户提供的可信文档中的机制,确保回答仅基于给定材料,消除幻觉。
|
||||
|
||||
## Core Mechanism
|
||||
NotebookLM 的知识库严格限制为用户上传的文档。AI 回答时:
|
||||
1. 仅检索用户文档中的相关段落
|
||||
2. 每个答案附带精确引文(指向原文)
|
||||
3. 用户可点击引文跳转到原文确认
|
||||
|
||||
## Why It Matters
|
||||
传统 LLM(GPT-4、Claude 等)依赖自身参数知识,存在幻觉风险。Source Grounding 通过强制锚定到可信文档,将错误率降至最低。
|
||||
|
||||
## Relationship to RAG
|
||||
RAG(Retrieval-Augmented Generation)也是一种 grounding 机制,但通常在开放域中检索。Source Grounding 是更严格的版本——仅限用户明确提供的文档。
|
||||
|
||||
## Connections
|
||||
- [[NotebookLM]] ← 典型实现
|
||||
- [[RAG]] ← 上位概念
|
||||
- [[引文追溯]] ← 机制保障
|
||||
Reference in New Issue
Block a user