传统知识库的问题
很多企业知识库依赖切片检索。切片太小,模型拿不到完整上下文;切片太大,召回不稳定;文档之间的关系也容易丢失。
长上下文带来的变化
GLM 5.3 支持 1M token 上下文,这让一些任务可以从“只看几个片段”升级为“读取一组完整资料”。
- 跨文档对比:比较多个版本、多个合同或多个产品说明。
- 长报告生成:读取完整背景后生成管理层摘要和行动清单。
- 复杂问答:回答需要结合多个章节、多个系统记录的问题。
- 知识校验:列出依据和不确定点,降低凭空回答风险。
仍然需要检索和权限
长上下文并不意味着不需要 RAG。更好的方式是:先用检索和权限系统选出相关资料,再把更完整的上下文交给模型处理。
The problem with traditional knowledge bases
Many enterprise knowledge bases rely on chunk retrieval. If chunks are too small, the model misses context; if they are too large, retrieval becomes unstable; relationships across documents can also disappear.
What long context changes
GLM 5.3 supports a 1M token context window, allowing some tasks to move from reading a few snippets to reading a fuller set of materials.
- Cross-document comparison across versions, contracts or product documents.
- Long report generation after reading fuller background materials.
- Complex Q&A that combines multiple sections and system records.
- Knowledge verification with evidence and uncertainty notes.
Retrieval and permissions still matter
Long context does not remove the need for RAG. A better pattern is to use retrieval and permissions to select relevant materials first, then pass richer context to the model.
想把企业文档改造成可问答知识库?Want to turn enterprise documents into a Q&A knowledge base?
咨询知识库方案Ask about knowledge workflows资料来源:Source: Z.AI GLM-5.3 Docs.