AITOPAI热报
How-toSelected

Milvus 3.0 教程：在线搜索应用中无缝更换 embedding 模型

Milvus@milvusio
Post
View the post
Why we picked it

Milvus 官方手把手教你换 embedding 模型不停机，旧向量照常能搜，还带 Recall@10 实测数据，做检索的可以直接照抄流程。

Milvus 发布教程讲解如何在不停机的搜索应用中更换 embedding 模型。核心做法是在同一 collection 中同时保留新旧两套向量：先新增向量字段和索引，用已有文档 ID 回填数据，再在同一批查询上评估两个模型的表现。文中 SciFact 示例里，从 BGE-small-en-v1.5 升级到 BGE-base-en-v1.5，300 条测试查询的 Recall@10 从 0.8362 提升到 0.8742。教程还包括逐步切换流量和稳定后删除旧字段的完整步骤。

Original · Milvus

To change the embedding model in a live search app, you need to build new document vectors while the old ones remain searchable. 𝗠𝗶𝗹𝘃𝘂𝘀 𝟯.𝟬 𝗹𝗲𝘁𝘀 𝘆𝗼𝘂 𝗸𝗲𝗲𝗽 𝗯𝗼𝘁𝗵 𝘀𝗲𝘁𝘀 𝗼𝗳 𝘃𝗲𝗰𝘁𝗼𝗿𝘀 𝗶𝗻 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝗰𝗼𝗹𝗹𝗲𝗰𝘁𝗶𝗼𝗻 𝗱𝘂𝗿𝗶𝗻𝗴 𝘁𝗵𝗲 𝘁𝗿𝗮𝗻𝘀𝗶𝘁𝗶𝗼𝗻. 𝗢𝘂𝗿 𝗻𝗲𝘄 𝘄𝗮𝗹𝗸𝘁𝗵𝗿𝗼𝘂𝗴𝗵 𝗰𝗼𝘃𝗲𝗿𝘀 𝘁𝗵𝗲 𝗽𝗿𝗼𝗰𝗲𝘀𝘀: 1. Add the new vector field and index. 2. Backfill it using existing document IDs, while keeping new and updated documents in sync. 3. Check that the vectors and index are ready, then evaluate both models on the same queries. 4. Shift traffic gradually. Pair each query model with the field containing its document vectors. 5. After the rollout is stable and old requests have finished, remove the old field. In the SciFact example, upgrading from BGE-small-en-v1.5 to BGE-base-en-v1.5 raised Recall @10 from 0.8362 to 0.8742 across 300 test queries. Your resul milvus.io/blog/milvus-3-… data, which is why the evaluation and rollback steps matter. Read the walkthrough: https://t.co/gR0JssgulR 💬 0 🔄 0 ❤️ 1 👀 95 ⚡

#Milvus#embedding#BGE#RAG#向量数据库

Keep reading

Recent stories with the same tag or section
PapersBuildersMulti-sourceSelected14:48
Milvus tests hyperbolic embeddings in image search rankings

The geometry matters less than the query type: hyperbolic scoring helps descriptive captions but hurts broad category queries where rare taxonomy terms dominate the gap. Teams running image search should benchmark Lorentz against cosine on their own query logs instead of assuming hyperbolic is uniformly better. Milvus supports hyperbolic ranking through a standard inner product, so no new infrastructure is required to test it.

#Hyper3-CLIP#Milvus#OpenAI CLIP#embeddings
Event
Milvus@milvusio6 sourcesSource
How-toBuildersSelected23:50
Zilliz tests Jev model as semantic cache gateway guard

Semantic caching usually forces a tradeoff between cost savings and serving wrong answers to slightly different queries; a cheap guard model shifts that balance toward accuracy. Engineers running RAG or agent workloads can plug this pattern into existing caches like GPTCache to cut LLM spend without degrading quality. Milvus and Zilliz have active developer communities across ASEAN, so the benchmark and blog are directly accessible to teams in the region.

#Milvus#Zilliz#Jev#semantic caching
Milvus@milvusioSource
How-toBuildersSelected01:06
Milvus Function Chain Reranks Search with XGBoost in One Call

Reranking inside the vector database itself removes the need for a separate reranking service and an extra network hop, which simplifies production search stacks. Builders running e-commerce, feed, or RAG systems can reuse the open-source demo to blend relevance with revenue-driving signals. It compares with the common pattern of exporting candidates from Milvus and reranking them in an external model server.

#Milvus#XGBoost#RAG#vector database
Milvus@milvusioSource
How-toBuildersSelected00:11
Milvus benchmark tests Jev against LLMs in agentic search

The benchmark suggests a practical routing rule: swap general-purpose LLM calls for Jev only on narrow, high-frequency decisions like stopping checks, and keep dedicated rerankers like Voyage rerank-3 for re-ranking. Teams running RAG pipelines can use the latency and cost figures to decide where Jev cuts spend without hurting recall. Full results are on Milvus's blog.

#Jev#Milvus#RAG#Voyage
Milvus@milvusioSource