Weaviate • vector database(@weaviate_io)

Better prompts won't fix a broken context pipeline. Reliable agentic RAG depends on controlling ...

8.5内容质量

TL;DR · AI 摘要

可靠代理式RAG需控制信息到达模型的方式，而非单纯优化提示词。Weaviate提出上下文工程的五大系统框架。

核心要点

  • 查询增强需重构用户输入以提升检索精度
  • 分块策略需平衡精度与上下文完整性的矛盾
  • 长期记忆应外部存储并定期维护

结构提纲

按章节快速跳转。

  1. §核心论点

    提示优化无法弥补上下文管道缺陷

  2. ·查询增强

    重构用户输入提升检索系统可用性

  3. ·检索优化

    分块策略需平衡精度与上下文完整性

  4. ·记忆管理

    长期记忆应外部存储并定期维护

  5. ·工具控制

    工具描述需精简以避免上下文污染

  6. ·代理协调

    代理需处理多环节错误传播问题

思维导图

用一张图看清主题之间的关系。

查看大纲文本（无障碍 / 无 JS 友好）
  • 上下文工程五大系统
    • 查询增强
      • 重构用户输入
    • 检索优化
      • 分块策略平衡
    • 记忆管理
      • 外部存储维护
    • 工具控制
      • 精简工具描述
    • 代理协调
      • 错误传播处理

金句 / Highlights

值得收藏与分享的关键句。

#RAG#向量数据库#上下文工程#Weaviate
打开原文

Weaviate AI Database on X: "Better prompts won't fix a broken context pipeline. Reliable agentic RAG depends on controlling 𝘄𝗵𝗮𝘁 𝗶𝗻𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻 𝗿𝗲𝗮𝗰𝗵𝗲𝘀 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹, 𝘄𝗵𝗲𝗻 𝗶𝘁 𝗮𝗿𝗿𝗶𝘃𝗲𝘀, 𝗮𝗻𝗱 𝗶𝗻 𝘄𝗵𝗮𝘁 𝗳𝗼𝗿𝗺. In @victorialslocum’s latest video, she breaks… / X

Weaviate AI Database

@weaviate_io

Better prompts won't fix a broken context pipeline. Reliable agentic RAG depends on controlling 𝘄𝗵𝗮𝘁 𝗶𝗻𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻 𝗿𝗲𝗮𝗰𝗵𝗲𝘀 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹, 𝘄𝗵𝗲𝗻 𝗶𝘁 𝗮𝗿𝗿𝗶𝘃𝗲𝘀, 𝗮𝗻𝗱 𝗶𝗻 𝘄𝗵𝗮𝘁 𝗳𝗼𝗿𝗺. In

@

victorialslocum

’s latest video, she breaks context engineering into 5 interconnected systems: 𝟭. 𝗤𝘂𝗲𝗿𝘆 𝗮𝘂𝗴𝗺𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻 User input is messy. Rewriting, expanding, or decomposing queries helps translate intent into something your retrieval system can actually use. No retrieval technique can compensate for misunderstood intent. 𝟮. 𝗥𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹 Chunking creates a real trade-off: small chunks improve precision but lose surrounding meaning, while large chunks add context but consume more of the window. Start with a simple baseline, then benchmark recursive or semantic chunking on representative data. 𝟯. 𝗠𝗲𝗺𝗼𝗿𝘆 Dumping complete conversation history into the prompt creates noise. Long-term memory should live outside the model and only be retrieved when relevant. It also needs maintenance: pruning stale information, merging duplicates, resolving contradictions, and deliberately forgetting what is no longer useful. 𝟰. 𝗧𝗼𝗼𝗹𝘀 Tool descriptions, argument formats, and tool selection all consume context. Exposing 100 poorly described tools doesn't make an agent more capable. It makes the relevant tool harder to find. 𝟱. 𝗔𝗴𝗲𝗻𝘁𝘀 Agents connect everything together by deciding what to retrieve, which tools to call, what to retain, and how to recover when a step fails. But they also compound upstream errors. Wrong chunks, stale memories, and malformed tool calls propagate through the entire pipeline. The model is only one part of the system. Dependable agentic RAG comes from engineering how context flows around it. Watch here:

youtu.be/sUoNoaatRvU?si…

$

00:00

/$

4:02 PM · Aug 20, 2026

885

Views

4

1

6

3