向量数据库、Embedding 与语义搜索实践

Choosing the wrong approximate strategy in multi-vector retrieval causes a 6x drop in nDCG@10, exceeding model upgrade gains. Measure embedding space separability via MaxSim std dev: use TokenANN/MUVERA for high spread, LEMUR for low spread.

Meta’s Stella app v273 contains a fully functional but inactive on-device face recognition pipeline with three models, local vector DB, and notification components—technically ready but not enabled for regular users.

Senqi AI 使用 Milvus 向物理机器人注入长期语义记忆能力，解决真实世界任务中环境动态、任务无界、指令模糊和错误高成本等核心挑战。

Context defocus significantly impacts the Claude Code agent, with seven open-source tools effectively addressing this issue, reducing token consumption by 60-90%.

Current in-car media systems still rely on keyword-based search, but drivers naturally express needs through emotions, vibes, and intent—not terms. Sarvesh Talele’s project, built with Qdrant Edge, delivers a fully local, AI-powered media discovery system supporting voice, text, and mood-based semantic queries—no cloud needed, ensuring privacy-first, real-time experience.

RAG systems rely on embeddings that fail predictably: when queries use different terms than docs (e.g., ‘overtime’ vs ‘non-employee labor’), contain negations, or depend on exact IDs/codes, retrieval fails. The article argues enterprise reliability comes from upstream filtering (expert keywords, doc structure), not rerankers atop weak retrieval.

The article argues that rerankers—often treated as a ‘magic layer’ in RAG systems—still fail on core semantic challenges like negation, logical complementation, and domain-specific terms, while adding significant latency; experiments show that in some cases, pure embedding retrieval (e.g., text-embedding-3-large) outperforms or matches the ‘embedding + reranker’ combo.

Research reveals poor retrieval quality is the primary cause of high-fluency hallucinations in RAG systems—more convincing, confident, and wrong—while scaling models fails to fix the root issue.

Milvus 提出通过 compaction（段合并与物理删除）和 TTL（自动过期）两项内置机制，可显著降低向量数据库存储成本，尤其适用于会话数据、时效性 RAG 等有生命周期的数据场景。

Qdrant 指出多数“图RAG”实为套壳向量检索，而 @datagraphs 构建了真正融合图数据库与向量搜索的协同架构：通过 schema-first agent 调度并行图查与语义检索，实现可验证、低延迟、全栈可控的知识问答。

AWS 推出基于 Amazon Nova 多模态嵌入的视频语义搜索方案，可联合处理音视频、文本等多源信号，提升检索准确性与效率。

文章系统解析现代稀疏神经检索模型（如SPLADE++），对比关键词与稠密检索优劣，并展示其在Qdrant中的实践应用。

文章详解如何使用 Sentence Transformers 微调多模态嵌入与重排序模型，并以视觉文档检索任务为例展示显著性能提升。

Weaviate官方博客分享了大规模数据导入和向量化实践，重点介绍服务器端批处理、错误处理及媒体处理策略，解决速率限制和批量失败问题。

Weaviate 1.38版本发布，HFresh向量索引和MCP服务器达一般可用性，提升百亿级数据处理和LLM集成能力。

LlamaIndex推出liteparse-grpc，支持gRPC接口以增强服务间通信的文档处理能力。

使用Qdrant和Llama Index构建多租户搜索系统可实现数据隔离与高效扩展，无需管理多个集合。

持续学习的核心在于构建可扩展的记忆系统而非频繁重训练，Qdrant通过开源记忆层实现这一目标。

Milvus通过时间感知排序函数结合语义相似性提升搜索结果时效性，使用衰减模型动态调整文档相关性评分。

Milvus 提供三种衰减模型（指数、高斯、线性）优化语义搜索的新鲜度与相关性平衡。

Weaviate的托管版本现已在DigitalOcean上公开预览，提供基础设施管理服务。

Weaviate推出的Engram通过智能记忆管理解决AI代理学习时变笨的问题，采用异步处理实现记忆提取、冲突化解和结构化存储。