[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"pack-detail-vector-db-showdown-zh":3,"seo:pack:vector-db-showdown:zh":80},{"code":4,"message":5,"data":6},200,"操作成功",{"pack":7},{"slug":8,"icon":9,"tone":10,"status":11,"status_label":12,"title":13,"description":14,"items":15,"install_cmd":79},"vector-db-showdown","🧲","#6D28D9","stable","稳定","向量数据库横评","Chroma \u002F Weaviate \u002F Pinecone \u002F txtai \u002F Qdrant MCP + Cohere & Together embedding API — 按延迟 \u002F 成本 \u002F RAG 准度三选一。",[16,28,35,43,51,61,71],{"id":17,"uuid":18,"slug":19,"title":20,"description":21,"author_name":22,"view_count":23,"vote_count":24,"lang_type":25,"type":26,"type_label":27},262,"04367306-be4a-4f46-854d-dd2b4d0d429e","chroma-open-source-vector-database-ai-04367306","Chroma — Open-Source Vector Database for AI","Chroma is the open-source vector database and data infrastructure for AI applications. 27.1K+ GitHub stars. Simple 4-function API for embedding, storing, and querying documents. Supports Python, JavaS","AI Open Source",339,0,"en","skill","Skill",{"id":29,"uuid":30,"slug":31,"title":32,"description":33,"author_name":22,"view_count":34,"vote_count":24,"lang_type":25,"type":26,"type_label":27},266,"492f7d14-9545-43b7-8f9c-626f895b912e","weaviate-open-source-vector-database-scale-492f7d14","Weaviate — Open-Source Vector Database at Scale","Weaviate is an open-source vector database for semantic search at scale. 15.9K+ GitHub stars. Hybrid search (vector + BM25), built-in RAG, reranking, multi-tenancy, and horizontal scaling. BSD 3-Claus",332,{"id":36,"uuid":37,"slug":38,"title":39,"description":40,"author_name":41,"view_count":42,"vote_count":24,"lang_type":25,"type":26,"type_label":27},826,"0fc5f7e8-439d-414f-bdaf-b09e05e1af49","pinecone-managed-vector-database-production-ai-0fc5f7e8","Pinecone — Managed Vector Database for Production AI","Fully managed vector database for production AI search. Pinecone offers serverless scaling, hybrid search, metadata filtering, and enterprise security with zero infrastructure.","Pinecone",296,{"id":44,"uuid":45,"slug":46,"title":47,"description":48,"author_name":49,"view_count":50,"vote_count":24,"lang_type":25,"type":26,"type_label":27},285,"b732febc-d945-4500-92c6-f90049c36c56","txtai-all-one-embeddings-database-b732febc","txtai — All-in-One Embeddings Database","txtai is an all-in-one embeddings database for semantic search, LLM orchestration, and language model workflows. 10.4K+ GitHub stars. Vector search + SQL + RAG pipelines. Apache 2.0.","Script Depot",347,{"id":52,"uuid":53,"slug":54,"title":55,"description":56,"author_name":57,"view_count":58,"vote_count":24,"lang_type":25,"type":59,"type_label":60},610,"301ce58e-1c73-48a8-af58-dfa560e8a13c","qdrant-mcp-vector-search-engine-ai-agents-301ce58e","Qdrant MCP — Vector Search Engine for AI Agents","MCP server for Qdrant vector database. Gives AI agents the power to store and search embeddings for RAG, semantic search, and recommendation systems. 22,000+ stars on Qdrant.","MCP Hub",293,"mcp","MCP",{"id":62,"uuid":63,"slug":64,"title":65,"description":66,"author_name":67,"view_count":68,"vote_count":24,"lang_type":25,"type":69,"type_label":70},774,"dde04e91-9c33-4bbb-9cf6-6604b1681582","cohere-embed-multilingual-ai-embeddings-api-dde04e91","Cohere Embed — Multilingual AI Embeddings API","Generate high-quality multilingual embeddings for search and RAG. Cohere Embed v3 supports 100+ languages with specialized modes for documents, queries, and classification.","Cohere",284,"prompt","Prompt",{"id":72,"uuid":73,"slug":74,"title":75,"description":76,"author_name":77,"view_count":78,"vote_count":24,"lang_type":25,"type":26,"type_label":27},779,"da3bf81c-8928-41ba-b5c4-457355af582d","together-ai-embeddings-reranking-skill-agents-da3bf81c","Together AI Embeddings & Reranking Skill for Agents","Skill that teaches Claude Code Together AI's embeddings and reranking API. Covers dense vector generation, semantic search, RAG pipelines, and result reranking patterns.","Together AI",328,"tokrepo install pack\u002Fvector-db-showdown",{"pageType":81,"pageKey":8,"locale":82,"title":83,"metaDescription":84,"h1":13,"tldr":85,"bodyMarkdown":86,"faq":87,"schema":103,"internalLinks":113,"citations":126,"wordCount":139,"generatedAt":140},"pack","zh","向量数据库横评：Chroma \u002F Weaviate \u002F Pinecone \u002F Qdrant","七个向量数据库横向对比：Chroma \u002F Weaviate \u002F Pinecone \u002F txtai \u002F Qdrant MCP + Cohere & Together embedding API。按延迟 \u002F 成本 \u002F RAG 准度三选一。","七个向量 DB 选型横向对比，覆盖自建 \u002F 托管 \u002F embedding API 三层 —— 按延迟 \u002F 成本 \u002F RAG 准度选。TokRepo 一条命令装齐。","## 这个 pack 装了什么\n\n这个包把 **七个主流向量数据库选项** 摆在一起，让选型从一周评测压成 10 分钟决策。选型空间干净分三层：自建数据库、托管数据库、embedding API（很多顺带提供基础向量存储）。\n\n| # | 资产 | 层级 | 强在哪 |\n|---|---|---|---|\n| 1 | Chroma | 自建 | 单节点最简，本地原型最快 |\n| 2 | Weaviate | 自建 \u002F 托管 | 内置 BM25 + 向量混合检索 |\n| 3 | Pinecone | 仅托管 | 零运维扩缩，p95 可预测 |\n| 4 | txtai | 自建 | 一个 Python 库搞定 embed + 检索 |\n| 5 | Qdrant MCP | 自建 | 原生 MCP 服务器，agent 直接查 |\n| 6 | Cohere embeddings | API | 多语言质量最佳 |\n| 7 | Together embeddings | API | 批量任务 token 经济性最优 |\n\n这个 pack 故意只覆盖 **DB 层** —— 存向量 + 提供最近邻查询。上面的 retrieve-and-generate 编排（切片 \u002F 改写 \u002F 重排）放在 **RAG Pipelines** 包里，两个决策保持独立。\n\n## 为什么要慎重选\n\n多数团队头六个月跑在错的向量 DB 上，等出问题才发现。两个典型崩盘：\n\n- **从 Pinecone 起步，扩规模时账单爆炸**。Pinecone 按 pod 收费，1M 向量没问题，到 50M 开始贵。迁出要重 embed 一遍。\n- **从自建起步，扩规模时运维爆炸**。一台 Chroma 节点存了 30M 向量，发现单节点 ANN 索引不会优雅降级 —— 查询延迟一个季度从 50ms 飙到 800ms。\n\n慎重选意味着看三个轴：\n\n1. **你的向量量级下的召回率 vs 延迟**。ANN-Benchmarks 公布 recall@10 vs QPS 曲线；>10M 向量时 Qdrant 和 Pinecone 持续领先，Chroma 在 5M 以下没问题。\n2. **混合检索需求**。如果查询既要关键词过滤又要语义相似，Weaviate 的混合模式和 Qdrant 的 payload 过滤是分水岭 —— 给 Chroma 后挂 BM25 是噩梦。\n3. **运维姿态**。两人团队选 Pinecone 的「没服务器要看」就赢。已经在大规模跑 Postgres 的，**pgvector**（在 Postgres for Agents 包里）TCO 上经常打败所有选项。\n\n## 一条命令装齐\n\n```bash\n# 把整个 pack 装进当前项目\ntokrepo install pack\u002Fvector-db-showdown\n\n# 或只装单个\ntokrepo install qdrant-mcp\ntokrepo install chroma\n```\n\nTokRepo CLI 把自建选项的 Docker Compose 片段、托管 API 的环境变量模板、加载 10 万向量并测 p95 延迟的基准脚本一起装进项目。\n\n## 常见踩坑\n\n- **拿随机向量做基准**。随机向量的距离分布扁平 —— 所有索引看起来一样快。永远用领域真实 embedding 做基准（Wikipedia dump 是公开代理）。\n- **选错距离度量**。同样数据下 cosine \u002F dot product \u002F L2 排序不同。用 embedding 模型训练时的度量；OpenAI text-embedding-3 用 cosine，部分开源模型用 dot product。\n- **忽视 embedding 模型锁定**。用 Cohere embed 1 亿文档后想换 OpenAI，要全部重 embed。有些团队过渡期同时存两套模型 embedding 并行。\n- **把「向量 DB」当完整 RAG 方案**。这些工具都不做重排、查询改写、结果质量评估。要配 RAG Pipelines 包和 LLM Eval 包。\n- **低估过滤基数**。先按高基数字段（例如 user_id）预过滤再 ANN 检索，多数引擎召回率崩。要么后过滤，要么按用户建索引。\n\n## 这套不够用的时候\n\n数据集小（\u003C1M 向量）且已经有 Postgres，**pgvector** 在运维简单度上打败所有选项 —— 少一个服务要看。查询除了语义相似还要地理或图约束，看 **Neo4j + GDS** 或 **OpenSearch** k-NN —— tradeoff 不同但形状更干净。规模到 10 亿向量级，已经超出这个 pack 范围；找厂商谈 Vespa 或 Milvus 专用层。",[88,91,94,97,100],{"q":89,"a":90},"这些向量 DB 跑起来收费吗？","七个里五个免费：Chroma \u002F Weaviate \u002F txtai \u002F Qdrant MCP 服务器都是宽松开源，只付计算。Pinecone 仅托管，免费起步档（10 万向量）；Cohere 和 Together 按 embedding 调用百万 token 计费。pack 把开源和付费定价都标了，选的时候没意外。",{"q":92,"a":93},"跟 rag-pipelines pack 区别？","这个 pack 是存储层 —— 装向量的。rag-pipelines 是上面的编排：切片、查询改写、检索器集成、重排。向量 DB 选一次很少换；RAG 参数要持续调。分成两个 pack 可以独立升级。",{"q":95,"a":96},"Claude Code \u002F Cursor 能用吗？","能。Qdrant 自带 MCP 服务器，Claude Code 把向量库当工具用 —— 任何 agent prompt 里都能调 `qdrant.search()`。Chroma 和 Weaviate 有社区 MCP 服务器，覆盖在 Modern CLI Toolbelt 和 MCP Server Stack 包里。Cursor 用户走标准 MCP 集成。",{"q":98,"a":99},"生产里 Pinecone 跟 Qdrant 区别？","Pinecone 全托管 p95 可预测、零运维扩缩，但 5000 万向量后 pod 单价急涨。Qdrant 哪都能跑 —— 笔记本、K8s、Qdrant Cloud —— 在 ANN-Benchmarks 高 QPS 召回率持续领先。小团队预算够选 Pinecone；要自建或规模上对成本敏感选 Qdrant。",{"q":101,"a":102},"向量 DB 之间迁移的运维坑？","向量在不同 embedding 模型间不可移植，但同一个模型下在不同 DB 间可以。多数迁移崩在团队迁移过程中顺手改了 embedding 流水线（chunk 大小 \u002F 模型版本）。先冻结流水线，把 embedding 快照到 S3，迁 DB，验证抽查相同 ID 返回，然后再迭代流水线。",{"@context":104,"@type":105,"name":106,"description":107,"numberOfItems":108,"publisher":109},"https:\u002F\u002Fschema.org","CollectionPage","Vector DB Showdown","Compare Chroma, Weaviate, Pinecone, txtai, Qdrant MCP, plus Cohere and Together embedding APIs.",7,{"@type":110,"name":111,"url":112},"Organization","TokRepo","https:\u002F\u002Ftokrepo.com",[114,118,122],{"url":115,"anchor":116,"reason":117},"\u002Fzh\u002Fpacks\u002Frag-pipelines","RAG 流水线","DB 之上的检索+生成层",{"url":119,"anchor":120,"reason":121},"\u002Fzh\u002Fpacks\u002Fdocument-ai-pipeline","文档 AI 流水线","产生向量的数据入库层",{"url":123,"anchor":124,"reason":125},"\u002Fzh\u002Fpacks\u002Fpostgres-for-agents","Postgres for Agent","Postgres 内嵌 pgvector 的混合选项",[127,131,135],{"claim":128,"source_name":129,"source_url":130},"Chroma is an AI-native open-source embedding database with Python and JavaScript clients","chroma-core\u002Fchroma","https:\u002F\u002Fgithub.com\u002Fchroma-core\u002Fchroma",{"claim":132,"source_name":133,"source_url":134},"Qdrant offers an MCP server enabling AI agents to query and write to vector collections","qdrant\u002Fmcp-server-qdrant","https:\u002F\u002Fgithub.com\u002Fqdrant\u002Fmcp-server-qdrant",{"claim":136,"source_name":137,"source_url":138},"ANN-Benchmarks tracks recall vs queries-per-second for major vector index implementations","ann-benchmarks","https:\u002F\u002Fann-benchmarks.com",502,"2026-05-02T15:00:00Z"]