做电商搜索或企业级向量搜索的团队，终于不用在数据库外再搭一套规则引擎了——Boost Ranker 把业务逻辑直接塞进搜索里，省掉一个系统，建议直接试试。
Milvus 在 2.6.2 版本中推出了 Boost Ranker 功能，解决了电商和企业团队在使用向量搜索时遇到的痛点：语义匹配结果往往不符合业务需求（如优先展示有库存的商品）。传统做法是在向量数据库外构建第二套系统进行后处理，增加了维护成本。Boost Ranker 将业务规则直接集成到搜索过程中，通过 filter、weight 和 re-sort 三步操作，在一个搜索调用内完成排序，无需外部依赖、无需重建索引、无需维护第二套系统，且几乎没有延迟开销。
𝗙𝗼𝗿 𝗮 𝘄𝗵𝗶𝗹𝗲, 𝘄𝗲 𝗸𝗲𝗽𝘁 𝗵𝗲𝗮𝗿𝗶𝗻𝗴 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝗰𝗼𝗺𝗽𝗹𝗮𝗶𝗻𝘁 𝗳𝗿𝗼𝗺 𝗲-...
𝗙𝗼𝗿 𝗮 𝘄𝗵𝗶𝗹𝗲, 𝘄𝗲 𝗸𝗲𝗽𝘁 𝗵𝗲𝗮𝗿𝗶𝗻𝗴 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝗰𝗼𝗺𝗽𝗹𝗮𝗶𝗻𝘁 𝗳𝗿𝗼𝗺 𝗲-𝗰𝗼𝗺𝗺𝗲𝗿𝗰𝗲 𝗮𝗻𝗱 𝗲𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲 𝘁𝗲𝗮𝗺𝘀 𝘂𝘀𝗶𝗻𝗴 𝘃𝗲𝗰𝘁𝗼𝗿 𝘀𝗲𝗮𝗿𝗰𝗵 𝗮𝗻𝗱 𝗿𝗮𝗻𝗸𝗶𝗻𝗴 𝗶𝗻 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻: "Vector search gives us great semantic results, but they're not what our users need." For instance, they need items in stock on the top of results instead of perfect matches that are sold out. 𝗧𝗵𝗲𝗶𝗿 𝗳𝗶𝘅? 𝗕𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗮 𝘀𝗲𝗰𝗼𝗻𝗱 𝘀𝘆𝘀𝘁𝗲𝗺 𝗼𝗻 𝘁𝗼𝗽 𝗼𝗳 𝘁𝗵𝗲𝗶𝗿 𝘃𝗲𝗰𝘁𝗼𝗿 𝗱𝗮𝘁𝗮𝗯𝗮𝘀𝗲. They would build custom post-processing logic, its own rules, its own tests, its own ops overhead — all just to get results sorted in a certain way. 𝗧𝗵𝗮𝘁'𝘀 𝘁𝗵𝗲 𝗽𝗿𝗼𝗯𝗹𝗲𝗺 𝘄𝗲 𝘀𝗲𝘁 𝗼𝘂𝘁 𝘁𝗼 𝘀𝗼𝗹𝘃𝗲 𝘄𝗵𝗲𝗻 𝘄𝗲 𝘀𝗵𝗶𝗽𝗽𝗲𝗱 𝗕𝗼𝗼𝘀𝘁 𝗥𝗮𝗻𝗸𝗲𝗿 𝗶𝗻 𝗠𝗶𝗹𝘃𝘂𝘀 𝟮.𝟲.𝟮. Instead of a separate reranking layer bolted on after search, 𝗕𝗼𝗼𝘀𝘁 𝗥𝗮𝗻𝗸𝗲𝗿 𝗺𝗼𝘃𝗲𝘀 𝘆𝗼𝘂𝗿 𝗯𝘂𝘀𝗶𝗻𝗲𝘀𝘀 𝗿𝘂𝗹𝗲𝘀 𝗶𝗻𝘁𝗼 𝘁𝗵𝗲 𝘀𝗲𝗮𝗿𝗰𝗵 𝗶𝘁𝘀𝗲𝗹𝗳: • 𝗳𝗶𝗹𝘁𝗲𝗿 picks the candidates that match your conditions • 𝘄𝗲𝗶𝗴𝗵𝘁 multiplies their scores up or down • the full set 𝗿𝗲-𝘀𝗼𝗿𝘁𝘀 into your final Top-K 𝗜𝘁'𝘀 𝗻𝗼𝘄 𝗼𝗻𝗲 𝘀𝗲𝗮𝗿𝗰𝗵 𝗰𝗮𝗹𝗹 𝘄𝗶𝘁𝗵 𝗻𝗼 𝗲𝘅𝘁𝗲𝗿𝗻𝗮𝗹 𝗱𝗲𝗽𝗲𝗻𝗱𝗲𝗻𝗰𝗶𝗲𝘀, 𝗻𝗼 𝗶𝗻𝗱𝗲𝘅 𝗿𝗲𝗯𝘂𝗶𝗹𝗱, 𝗮𝗻𝗱 𝗻𝗼 𝘀𝗲𝗰𝗼𝗻𝗱 𝘀𝘆𝘀𝘁𝗲𝗺 𝘁𝗼 𝗺𝗮𝗶𝗻𝘁𝗮𝗶𝗻. 𝗜𝗻 𝗮𝗱𝗱𝗶𝘁𝗶𝗼𝗻 𝘁𝗼 𝘁𝗵𝗮𝘁: • 𝗧𝗵𝗲𝗿𝗲'𝘀 𝗻𝗼 𝗺𝗲𝗮𝗻𝗶𝗻𝗴𝗳𝘂𝗹 𝗹𝗮𝘁𝗲𝗻𝗰𝘆 𝗰𝗼𝘀𝘁 Boost Ranker runs on the already-retrieved candidate set, not the full dataset. The operations are simple filter-and-multiply, so the overhead is negligible compared to the search itself. • 𝗥𝘂𝗹𝗲𝘀 𝗮𝗿𝗲 milvus.io/blog/milvus-bo… �𝗴𝘂𝗿𝗲 𝗮𝗻𝗱 𝘄𝗼𝗿𝗸 𝗮𝗹𝗼𝗻𝗴𝘀𝗶𝗱𝗲 𝗲𝘃𝗲𝗿𝘆𝘁𝗵𝗶𝗻𝗴 𝗲𝗹𝘀𝗲 Need to boost official documents, demote stale content, and add a diversity factor? Stack multiple Function objects and combine them with FunctionScore. boost_mode controls how each rule interacts with the original score; function_mode controls how rules interact with each other. It all happens inside a single search call — nothing separate to wire up or maintain. Full breakdown → https://t.co/cgqPHX3Jio 💬 0 🔄 0 ❤️ 0 👀 12 ⚡