Back to home

Compare

Comparing: RAG Knowledge Bases Enter the Engineering Phase: Permissions and Data Quality Matter More Than & RAG 知识库进入工程化阶段 — 权限与资料质量比模型更关键

AEN
RAGDifyRAGFlow·

RAG Knowledge Bases Enter the Engineering Phase: Permissions and Data Quality Matter More Than

The mainstream solution for enterprise knowledge bases is still RAG, but what usually determines answer quality is not the strength of the model. It is the quality of the source material, retrieval accuracy, and permission design. The pragmatic path outlined in the source article is to validate the system with 50–200 core documents first, then expand based on actual questions. This is far more worth watching than building a complex system from day one.

What this is

RAG (Retrieval-Augmented Generation) works by first splitting documents into chunks and building vector and keyword indexes. When a user asks a question, the system retrieves the relevant source text and passes it to a large language model to generate an answer. Because it does not write the source material into the model’s parameters, a knowledge base can remain current and traceable, while making it easier to control the evidence behind each response.

A complete pipeline includes document ingestion, cleaning and chunking, hybrid retrieval, permission filtering, reranking, generation, and updating. Chunking typically starts at 300–800 Chinese characters, with approximately 10%–20% overlap between adjacent chunks. During retrieval, vector search and keyword search run simultaneously, narrowing the candidate results to three to five passages.

The article recommends that enterprises initially use platforms such as Dify, FastGPT, or RAGFlow to process a small set of documents. When there is a genuine need for deep customization, they can then build their own systems with LlamaIndex or LangChain. Vector databases can include Milvus, Qdrant, or the pgvector extension for Postgres. However, a vector database is only part of the storage layer, and replacing it on its own will not usually improve results significantly.

Industry view

We believe enterprise knowledge bases are moving from the “model demo” stage into the “engineering governance” stage. Every answer must cite its original source, and when retrieval finds nothing, the system must clearly state that the information is not in the available material. Otherwise, fluent responses can easily conceal errors in the underlying documents. More importantly, permission filtering must happen before retrieval, not after generation. Otherwise, content the user is not authorized to access may already have entered the model’s context.

Yet this approach is not a low-cost shortcut. Document parsing, permission synchronization, version replacement, and evaluation all require sustained investment. Open-source platforms such as RAGFlow and Dify have not eliminated the operational burden. If teams only append content and never remove old chunks, outdated policies and newer versions will coexist. Cross-document summarization also cannot simply rely on a single retrieval pass; it requires a separately built summary or topic index.

Impact on regular people

For enterprise IT: The priority in a knowledge base project should shift from procuring a larger model to organizing source material, maintaining permissions, and establishing an evaluation set of 30–50 real-world questions. Without a measurable question set, teams can easily make technology decisions based on the polish of a demo.

For individual professionals: Policies, APIs, and troubleshooting materials will become easier to find faster, but users will still need to verify citations. For answers involving compensation, customers, or technical details, users should confirm them against the original files.

For the consumer market: Enterprise assistants with source citations will gradually become standard in office software and customer service. At the same time, the risks of storing document copies in the cloud, licensing terms, and data leaving the organization will become practical constraints when individuals evaluate personal editions of these tools.

Source: juejin.cn
BZH
RAGDifyRAGFlow·

RAG 知识库进入工程化阶段 — 权限与资料质量比模型更关键

企业知识库的主流方案仍是 RAG,但决定回答质量的通常不是模型强弱,而是资料质量、检索准确率和权限设计。源文章给出的务实路径,是先用50—200篇核心文档验证,再按问题扩展,这比一开始自建复杂系统更值得关注。

这是什么

RAG(检索增强生成)指先把文档切成片段并建立向量、关键词索引,用户提问时先找回相关原文,再交给大模型据此回答。它不把资料写入模型参数,因此知识库可以更新、可追溯,也更容易控制回答依据。

一套完整链路包括资料接入、清洗切块、混合检索、权限过滤、重排序、生成和更新。切块通常先从300—800字试起,相邻片段保留约10%—20%的重叠;检索时同时使用向量搜索和关键词搜索,并将候选结果压缩至3—5段。

文章建议企业先采用 Dify、FastGPT、RAGFlow 等平台处理小范围文档;确有深度定制需求时,再使用 LlamaIndex 或 LangChain 自建。向量库可选 Milvus、Qdrant 或 Postgres 的 pgvector 扩展,但它只是存储环节,单独更换通常不能显著改善效果。

行业怎么看

我们认为,企业知识库正在从“模型演示”进入“工程治理”阶段。回答必须附原文来源,检索不到就明确说资料里没有;否则流畅表达很容易掩盖资料错误。更重要的是,权限过滤应发生在检索之前,而不是生成之后,否则无权限内容可能已经进入模型的上下文。

但这套方案并非低成本捷径。文档解析、权限同步、版本替换和评测都需要持续投入,RAGFlow、Dify 等开源平台也没有消除运维负担。只追加而不删除旧片段,会让过期制度与新版本同时存在;跨文档总结也不能简单依赖一次检索,需要单独建设摘要或主题索引。

对普通人的影响

对企业 IT:知识库项目的重点应从采购更大模型转向整理资料、维护权限和建立30—50道真实问题的评测集。没有可衡量的问题集,团队很容易凭演示效果做技术决策。

对个人职场:制度、接口和排错资料会更快被找到,但使用者也需要核查引用。涉及薪酬、客户或技术细节时,答案仍应回到原始文件确认。

对消费市场:带来源的企业助手会逐步进入办公软件和客服场景;与此同时,文档副本云端保存、许可条款及数据出域风险,也会在采购个人版工具时成为实际约束。

Source: juejin.cn