A vector search can find the right passage but still miss the surrounding instructions. 𝗜𝗻 ...
TL;DR · AI 摘要
Milvus 3.0的StructArray技术通过单次查询获取匹配块及其上下文,提升RAG系统效率。
核心要点
- StructArray将文档存储为单个实体,避免额外查询
- output_fields参数可同时获取匹配块及相邻块
- 网络研讨会演示了汽车轮毂指南的完整应用案例
结构提纲
按章节快速跳转。
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)
- StructArray技术
- 文档存储方式
- 单实体存储
- 查询优化
- output_fields参数
- 单次查询
- 应用场景
- RAG系统优化
金句 / Highlights
值得收藏与分享的关键句。
StructArray stores the whole document as one entity, with chunks in their original order.
By requesting the chunk text in output_fields, the app can pull chunks 0, 1, and 2 from the same response.
Check the full webinar on YouTube for schema and queries.
Milvus on X: "A vector search can find the right passage but still miss the surrounding instructions. 𝗜𝗻 𝗼𝘂𝗿 𝗠𝗶𝗹𝘃𝘂𝘀 𝟯.𝟬 𝘄𝗲𝗯𝗶𝗻𝗮𝗿, 𝗥𝗼𝗵𝗶𝘁 𝗡𝗶𝗷𝗵𝗮𝘄𝗮𝗻 𝘀𝗵𝗼𝘄𝘀 𝗵𝗼𝘄 𝘁𝗼 𝘂𝘀𝗲 𝗦𝘁𝗿𝘂𝗰𝘁𝗔𝗿𝗿𝗮𝘆 𝘁𝗼 𝗳𝗲𝘁𝗰𝗵 𝗮 𝗺𝗮𝘁𝗰𝗵𝗶𝗻𝗴 𝗰𝗵𝘂𝗻𝗸 𝗮𝗻𝗱 𝗶𝘁𝘀 𝗻𝗲𝗶… / X
@milvusio
A vector search can find the right passage but still miss the surrounding instructions. 𝗜𝗻 𝗼𝘂𝗿 𝗠𝗶𝗹𝘃𝘂𝘀 𝟯.𝟬 𝘄𝗲𝗯𝗶𝗻𝗮𝗿, 𝗥𝗼𝗵𝗶𝘁 𝗡𝗶𝗷𝗵𝗮𝘄𝗮𝗻 𝘀𝗵𝗼𝘄𝘀 𝗵𝗼𝘄 𝘁𝗼 𝘂𝘀𝗲 𝗦𝘁𝗿𝘂𝗰𝘁𝗔𝗿𝗿𝗮𝘆 𝘁𝗼 𝗳𝗲𝘁𝗰𝗵 𝗮 𝗺𝗮𝘁𝗰𝗵𝗶𝗻𝗴 𝗰𝗵𝘂𝗻𝗸 𝗮𝗻𝗱 𝗶𝘁𝘀 𝗻𝗲𝗶𝗴𝗵𝗯𝗼𝘂𝗿𝘀 𝗶𝗻 𝗮 𝘀𝗶𝗻𝗴𝗹𝗲 𝗰𝗮𝗹𝗹. When a RAG pipeline stores each chunk as a separate row, getting the surrounding text can take an extra query. StructArray stores the whole document as one entity, with chunks in their original order. Each chunk has its own text and vector, so you can still search at the passage level. His demo uses a guide to buying car rims, split into eight chunks. The search hits the paragraph about bolt patterns at offset=1. By requesting the chunk text in output_fields, the app can pull chunks 0, 1, and 2 from the same response and pass them to the LLM in reading order. Check the full webinar on YouTube for schema and queries.
youtube.com/watch?v=OT9XKJ…
#Milvus
#RAG
$
00:00
/$
3:20 PM · Nov 9, 2026
·
213
Views
1
2