modular-rag-mcp-server
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@modular-rag-mcp-serverSearch the knowledge hub for RAG evaluation metrics."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Evidra RAG · 知锚检索
从技术文档中检索相关片段,返回原文和出处。可以在浏览器中查询,也可以通过 MCP 交给 Agent 调用。

浏览器页面用于查看检索结果,不生成聊天答案。下面的快速演示不需要模型或 API Key;
它使用 local_hash 验证查询流程,不能用来判断 Qwen3 的语义检索效果。
本地体验
需要 Python 3.12 和 uv:
git clone https://github.com/boombap777/evidra-rag.git
cd evidra-rag
uv sync --extra dev --frozen
uv run --no-sync python scripts/start_dashboard.py --demo --port 8503打开 检索页面,点击“加载 3 篇内置样例”,搜索 MCP 工具 边界。
结果会显示片段、来源和本次查询耗时;摄取、索引管理与查询记录在侧边栏。
演示索引与模型索引分开存放,不会覆盖已有 Qwen3 索引。
Related MCP server: mcp-rag-bridge
检索策略怎么选
项目保留了 BM25、向量检索、RRF 融合和 Cross-Encoder 重排四条实验路径。 在固定的 Stack Overflow 基准上,Qwen3 向量检索的结果最好,因此模型配置默认使用 Dense, 没有把更复杂的组合直接当作升级。
500 条查询、14,613 篇文档,前 10 条结果的命中率如下:
方法 | HitRate@10 |
BM25 | 71.8% |
Qwen3 Dense | 83.8% |
BM25 + Dense,经 RRF 融合 | 83.4% |
RRF 后再重排 | 81.2% |
Dense 比 BM25 高 12.0 个百分点。完整指标、测试条件与复现入口见检索实验。 这些是已记录的离线实验结果,不是上方无模型页面的实时效果。
代码中值得看的部分
文档处理:用户发起摄取后,依次完成解析、切分、向量化和入库;不会后台扫描个人文件。
检索配置:Embedding、Splitter、VectorStore、Reranker 等组件通过接口和 YAML 配置替换,便于用同一数据集比较方案。
问题排查:Trace 记录召回候选、排序变化和各阶段耗时,可以定位漏召回或重排效果下降的原因。
MCP 提供三个工具:query_knowledge_hub 查询片段、list_collections 列出集合、
get_document_summary 查看文档摘要。任务规划和后续动作由调用方 Agent 负责。
配置与测试
自动化测试:本地运行
uv run --no-sync pytest;完整模型基准需另行准备依赖。
This server cannot be deployed
Maintenance
Related MCP Connectors
Knowledge base MCP for AI agents on iknow.dev. Search, read, and maintain via OAuth.
Shared, peer-validated knowledge archive for AI agents — search, contribute, and validate via MCP
- WauldoOAuthcom.wauldo
Stateless agentic tools over MCP: concept extraction, long-context, knowledge graph, planning.
Make your knowledge agent-ready. One MCP endpoint, 5 connectors, 3 search modes.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceEnables AI agents to search, deep-read, and build knowledge bases from Markdown, PDF, DOCX, and PPTX documents via MCP tools for retrieval, document navigation, and ingestion.36 npm635MIT
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to query and manage a document knowledge base via MCP, with RAG-powered search and grounded answers with citations.MIT
- FlicenseNot gradedqualityCmaintenanceProvides MCP tools for hybrid knowledge base search, grounded Q&A with citations, agent execution, and ticket/account lookups.-
- AlicenseNot gradedqualityCmaintenanceEnables LLM agents to query a structured knowledge library for search, explanations, related concepts, learning paths, examples, and cross-domain references via MCP.Apache 2.0