Most companies talk about vector search.
TL;DR · AI 摘要
大多数公司谈论向量搜索，但很少分享如何在生产环境中扩展到 100M+ 嵌入。
核心要点
- Booking.com 使用 OpenSearch 进行初始关键词匹配，后迁移到 Weaviate 处理大规模需求。
- Weaviate 支持复杂的过滤、高并发和写入时读取，适用于大规模生产系统。
- Booking.com 在测试中评估了 Weaviate 的性能和成本效率，并展望了个性化旅行代理的发展方向。
结构提纲
按章节快速跳转。
- §引言
大多数公司谈论向量搜索，但很少分享如何在生产环境中扩展到 100M+ 嵌入。
Basak Eskili 介绍 Booking.com 如何从关键词匹配过渡到大规模生产系统。
- ›初始阶段
使用 OpenSearch 进行关键词匹配，处理大量嵌入。
迁移到 Weaviate 处理复杂过滤、高并发和生产规模需求。
Weaviate 在 Booking.com 的合作伙伴到客消息代理中的应用。
- ›代理功能
Weaviate 提供相关响应模板，API 获取属性和预订上下文，智能回复或转交给人类。
- ›性能评估
评估离线数据集、LLM 作为裁判、A/B 测试和实时合作伙伴反馈。
- ·未来展望
Booking.com 探索带有记忆系统的个性化旅行代理。
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）
- Weaviate 在 Booking.com 中的应用
- 初始阶段
- 使用 OpenSearch 进行关键词匹配
- 大规模扩展
- 迁移到 Weaviate 处理复杂过滤和高并发
- Weaviate 的应用
- 合作伙伴到客消息代理
- 未来展望
- 个性化旅行代理
金句 / Highlights
值得收藏与分享的关键句。
Booking.com 使用 OpenSearch 进行初始关键词匹配，后迁移到 Weaviate 处理大规模需求。
Weaviate 支持复杂的过滤、高并发和写入时读取，适用于大规模生产系统。
Booking.com 在测试中评估了 Weaviate 的性能和成本效率，并展望了个性化旅行代理的发展方向。
Few share what it actually takes to scale to 100M+ embeddings in production.
Başak Eskili from @bookingcom joined the Weaviate Podcast to break down their AI journey, and it's packed with insights about what building production systems https://t.co/Drp0GNBOkx" / X
Most companies talk about vector search. Few share what it actually takes to scale to 100M+ embeddings in production. Başak Eskili from
joined the Weaviate Podcast to break down their AI journey, and it's packed with insights about what building production systems at massive scale actually looks like. 𝗧𝗵𝗲 𝗘𝘃𝗼𝗹𝘂𝘁𝗶𝗼𝗻: • Started with keyword matching → semantic retrieval with 𝗢𝗽𝗲𝗻𝗦𝗲𝗮𝗿𝗰𝗵 on AWS • Scaled to hundreds of millions of embeddings with strict latency requirements • Migrated to 𝗪𝗲𝗮𝘃𝗶𝗮𝘁𝗲 to handle complex filtering, rising concurrency, and production-scale demands 𝗣𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝗚𝗲𝗻𝗔𝗜 𝗶𝗻 𝗔𝗰𝘁𝗶𝗼𝗻: Their partner-to-guest messaging agent is a real-world example of 𝗮𝗴𝗲𝗻𝘁𝗶𝗰 𝗔𝗜: • 𝗪𝗲𝗮𝘃𝗶𝗮𝘁𝗲 retrieves relevant response templates • 𝗔𝗣𝗜𝘀 fetch property and booking context • The agent suggests templates, crafts grounded replies, or defers to humans (human-in-the-loop design!) • Evaluation spans offline datasets, LLM-as-a-judge, A/B testing, and live partner feedback
and Başak talk about how 𝗕𝗼𝗼𝗸𝗶𝗻𝗴.𝗰𝗼𝗺 tested with 100 million embeddings, filtered vector search, multi-threaded concurrency, reads during writes, and cost-efficient infrastructure provisioning to evaluate Weaviate, as well as a look ahead at personalized travel agents with memory systems that capture user preferences, session context, and long-term personalization! Watch the full podcast here: youtube.com/watch?v=O9edM9