返回首页

对比阅读

对比阅读:RAG Retrieval Quality Sets the Ceiling for Model Accuracy 与 RAG 让模型先查资料再回答,但检索准确度决定答案上限

AEN
RAGEmbedding ModelsDeepSeek·

RAG Retrieval Quality Sets the Ceiling for Model Accuracy

Andrew Ng once offered a widely cited assessment: roughly 90% of Agent projects do not fail because of model capability, but because of implementation details. RAG (retrieval-augmented generation, which retrieves information before the model answers) is one of the most fundamental components in that stack: it does not make models smarter, but helps them find more relevant material before responding.

Connecting a vector database, however, is not enough to get it right. Whether documents are chunked sensibly, relevant material is recalled accurately, results are ranked reliably, and the model uses the retrieved content faithfully all determine the final answer. If retrieval fails, even a fluent response may simply express an error with greater confidence.

What This Is

The basic RAG workflow has two stages: building the knowledge base and querying it. During indexing, an embedding model converts each passage into a sequence of numbers, or vectors; passages with similar meanings sit closer together in this “semantic space.” At query time, the system converts the question into a vector, identifies the most relevant passages, places them in the model’s context, and generates an answer based on them.

It typically involves three stages: chunking, which splits long documents into manageable passages for retrieval; recall, which retrieves semantically relevant candidates; and ranking, which places the content most likely to contain the answer first. “Semantic retrieval” therefore compares not merely whether two passages use the same words, but whether they convey similar meanings.

Industry View

We believe RAG has moved from a research concept to infrastructure for enterprise AI projects. Its value lies not in technological showmanship, but in turning internal documents, policies, product materials, and historical records into searchable knowledge entry points. For enterprise use cases where knowledge changes frequently and answers must be traceable, it is more practical than simply expanding a model’s context window.

Nevertheless, the objections are equally clear: retrieval is not fact-checking in itself. If chunking breaks semantic coherence, recall misses critical material, or ranking puts secondary information first, the system may produce an answer with citations that does not actually address the question. More problematic still, embedding models and chat models have different roles. Replacing either model while retaining the old index often creates a mismatch between the vector space and the new model.

Moreover, RAG adds the costs of document governance, access control, index updates, and result evaluation. It can improve answers, but it cannot automatically eliminate outdated materials, contradictory information, or unreliable sources. Without clear ownership and maintenance mechanisms for their data, enterprises can easily turn a knowledge base into a larger repository of confidently wrong answers.

Impact on Regular People

For enterprise IT: When companies deploy AI customer service, internal assistants, or compliance Q&A systems, the real workload shifts from training models to organizing information, configuring permissions, and continuously evaluating retrieval performance. RAG lowers the barrier to using general-purpose models, but it cannot replace foundational data governance.

For professionals: AI assistants connected to company information may increasingly explain workflows, contracts, and historical projects more accurately. Whether they are worth using depends not only on how quickly they respond, but also on whether they provide sources and recognize when the available material does not contain an answer.

For the consumer market: Products with retrieval capabilities will gradually reduce instances of models confidently fabricating answers, but consumers will still need to verify that citations correspond to the original passages and that the underlying material is current. Wider adoption does not make answers exempt from scrutiny.

来源: juejin.cn
BZH
RAG嵌入模型DeepSeek·

RAG 让模型先查资料再回答,但检索准确度决定答案上限

Andrew Ng 曾给出一个常被引用的判断:约 90% 的 Agent 项目并非卡在模型能力,而是卡在落地细节。RAG(检索增强生成,先检索资料再让模型回答)正是其中最典型的基础环节:它不是让模型变聪明,而是让模型在回答前找到更可能相关的内容。

但这件事并不是“接上向量数据库就完成”。资料是否切得合理、召回是否准确、排序是否可靠,以及模型是否忠实使用检索结果,共同决定最终答案。检索错了,模型答得再流畅,也可能只是把错误说得更自信。

这是什么

RAG 的基本流程分为建库和查询两步。建库时,嵌入模型把每段文字转换为一串数字,也就是向量;意思相近的文字,向量在“意义空间”里也更接近。查询时,系统把问题转成向量,找出最相关的几段内容,放进模型的上下文中,再据此生成回答。

它通常包含三个环节:切块,把长文档拆成适合检索的小段;召回,找回语义相关的候选内容;排序,把最可能包含答案的内容放到前面。所谓“语义检索”,就是不再只比较字面词是否相同,而是比较两段内容表达的意思是否接近。

行业怎么看

我们认为,RAG 已经从研究概念变成企业 AI 项目的基础设施,其价值不在于炫技,而在于把内部文档、制度、产品资料和历史记录变成可查询的知识入口。对于知识更新频繁、答案必须可追溯的企业场景,它比单纯扩大模型上下文更实际。

不过,反对意见同样明确:检索不是事实核验本身。只要切块破坏语义、召回遗漏关键材料,或者排序把次要信息放在前面,系统就可能产生“有出处但答非所问”的结果。更棘手的是,嵌入模型和聊天模型职责不同,替换其中一个模型却沿用旧索引,往往会造成向量空间不匹配。

此外,RAG 会增加文档治理、权限控制、索引更新和结果评估的成本。它能改善答案,却不能自动消除过期资料、相互矛盾的信息,以及来源本身不可靠的问题。企业若没有维护资料的责任机制,所谓知识库很容易变成规模更大的错误答案库。

对普通人的影响

对企业 IT:企业部署 AI 客服、内部助手或合规问答时,真正的工作量会从训练模型转向整理资料、设置权限和持续评估检索效果。RAG 能降低调用通用模型的门槛,但不能替代基础数据治理。

对个人职场:与公司资料连接起来的 AI 助手,未来可能更准确地解释流程、合同和历史项目。是否值得使用,关键不只看模型回答得多快,还要看它能否给出来源,以及何时承认“资料里没有”。

对消费市场:带检索能力的产品会逐步减少“一本正经地编答案”的情况,但消费者仍需检查引用是否对应原文、资料是否已经更新。功能普及不意味着答案可以免于核实。

来源: juejin.cn