RAG · 答案背后的五站链路RAG · The Five Stations Behind an Answer
retrieve → filter → load → generate → cite
检索增强生成(Retrieval-Augmented
Generation,RAG)先从外部知识源找证据,再把证据放入上下文辅助生成;它不训练新参数,却能更新知识并留下来源链。Retrieval-augmented generation finds evidence in an external source and
loads it into context before generation. It updates available
knowledge without retraining parameters and can preserve a source
trail.
01 · 一句答案从哪里来01 · Where did that answer come from?
如果答案能说出最新退货期限,却看不到它经过哪些站,你敢直接采用吗?If an answer knows the latest return window but hides every station
it passed through, would you trust it?
R→A
02五站总控台:逐站断开链路Five-Station Console: Break the Chain One Station at a Time
总门面chain facade
可用证据Usable evidence
1
片段passage
答案状态Answer state
30天
supported
证据链Evidence chain
5/5
站点连通stations connected
03五站不是一个黑盒Five Stations, Not One Black Box
输入 · 证据 · 输出input · evidence · output
RETRIEVE
检索扩大候选Retrieval finds candidates
按查询从外部知识库召回可能相关的片段。Recall possibly relevant passages from an external knowledge
base.
FILTER + LOAD
筛选并装入上下文Filter and load context
去掉旧版与无关项,只把合格证据交给生成阶段。Remove stale or irrelevant items and pass only eligible evidence
into context.
GENERATE
生成使用证据Generation uses evidence
模型依据当前问题和已装载片段组织回答。The model composes an answer from the question and loaded
passages.
CITE
引用连接主张Citation links the claim
把“30
天”定位到具体版本、条款和片段,而非只列来源。Tie “30 days” to a specific version, clause, and passage
instead of merely listing a source.
04最小机制:后站依赖前站产物Minimal Mechanism: Each Station Depends on the Previous
Output
A = G(q, C(E))
E = filter(retrieve(q)) A = generate(q, load(E))
引用 c(A, E) 只有在答案主张能定位到证据 E
时才成立。关掉前站,后站不会凭空补齐缺失产物。Citation c(A, E) is valid only when an answer claim resolves to
evidence E. A downstream station cannot invent a missing
upstream output.
05边界实验:答案对了,却断了来源Boundary Test: The Answer Looks Right but Loses Its Source
主动制造失败Create a failure
只关闭“引用”站Disable only the citation station
答案仍会显示“30
天”,但读者无法检查版本和条款。内容看似正确,不等于证据链完整。The answer still says “30 days,” but the reader cannot
inspect its version or clause. A plausible answer is not a
complete evidence chain.
尚未执行。当前五站全部连通。Not run yet. All five stations are connected.
✗ RAG 会把新知识训练进参数。RAG trains new knowledge into parameters.
→ ✓ 它在推理时检索并装载外部证据。It retrieves and loads external evidence at inference
time.
✗ 只要检索到了,回答就一定会用。If a passage was retrieved, the answer must have used
it.
→ ✓ 筛选、装载和生成仍可能各自断链。Filtering, loading, and generation can each break the
chain.
✗ 答案正确就不需要引用。A correct answer needs no citation.
→ ✓ 引用让主张可定位、可复核。Citation makes the claim locatable and checkable.
06 · 一句话带走06 · One line to keep
RAG
是“检索→筛选→装载→生成→引用”的可检查链路;任何一站断开,答案内容或证据归因都会发生可见变化。RAG is an inspectable retrieve → filter → load → generate →
cite chain; breaking any station visibly changes the answer or its
attribution.